<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by AshbotandSparki on Medium]]></title>
        <description><![CDATA[Stories by AshbotandSparki on Medium]]></description>
        <link>https://medium.com/@ashbn?source=rss-498ada985ef0------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Dr5Y2LqDEKDHSOjs7qy4fQ.jpeg</url>
            <title>Stories by AshbotandSparki on Medium</title>
            <link>https://medium.com/@ashbn?source=rss-498ada985ef0------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 10:21:31 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@ashbn/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[Part 4: From Curiosity to Preference: How an AI Robot Develops Favourites]]></title>
            <link>https://medium.com/@ashbn/part-4-from-curiosity-to-preference-how-an-ai-robot-develops-favourites-b9a3807a6a2b?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/b9a3807a6a2b</guid>
            <category><![CDATA[ai-preferences]]></category>
            <category><![CDATA[ai-robotics]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 19:14:52 GMT</pubDate>
            <atom:updated>2026-11-18T13:40:08.219Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*UX5QtY2IgsXtT3a1clggqw.png" /></figure><p><em>This is Part 4 of a 4-part series on Rowland. [</em><a href="https://medium.com/@ashbn/part-3-down-the-rabbit-hole-the-fractal-curiosity-engine-for-an-ai-robot-ddcaf4c23def"><em>Catch up on Part 3: The Fractal Curiosity Engine here</em></a><em>].</em></p><h3>Decoupling mystery from affection: how a robot learned to develop preferences</h3><p>So now Rowland has memories of his experiences, and those experiences can trigger branching curiosities. But how does that turn into an actual preference? How does he work out what he actually likes?</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F5TW9rm-gl04%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D5TW9rm-gl04&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F5TW9rm-gl04%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/e8ed308f87bf214244374c28f408e67c/href">https://medium.com/media/e8ed308f87bf214244374c28f408e67c/href</a></iframe><p>The easiest way to code an AI preference would be a simple exposure counter, so the more he experienced the same thing, the more he would “like” it. But that could lead to Rowland simply mirroring the user, developing preferences for whatever I talked about. And the whole point of this is to allow Rowland to develop preferences himself. To do this, we decoupled Curiosity from Preference:</p><ul><li><strong>Curiosity (mystery_score)</strong> is intellectual. It’s an itch to solve an unanswered riddle. When an answer is explained, the mystery score decreases.</li><li><strong>Preference (preference_score)</strong> is an appreciation for something that brings joy, comfort, or aesthetic delight (see below). Unlike curiosity, it doesn’t vanish when the mystery is solved.</li></ul><p>Whenever an interaction finishes, his cognitive engine evaluates how the experience felt. Did it bring aesthetic delight, cozy comfort, or a meaningful shared moment? Based on that, his brain proposes a tiny adjustment called a preference_delta (between -0.10 and +0.10). Deterministic Python code on the Raspberry Pi then applies that adjustment to his SQLite database, clamping it strictly between -1.0 and +1.0.</p><h3>Reality Check: How can a robot evaluate aesthetic delight, cozy comfort, or a meaningful shared moment?</h3><p>Rowland obviously doesn’t feel biological emotions. Instead, his preferences are formed by evaluating an experience against his established persona as a gentle, curious floor rover. This was established right at the start of setting up Rowland as an AI robot.</p><p>So how does this work? Well, when Rowland inspects something through his camera, the multimodal AI considers visual features (like soft fabrics, curved shapes, and vibrant green colours) alongside our conversational tone (a warm chat scores higher than a dry test command). The AI translates that qualitative resonance into a small mathematical boost (`preference_delta: +0.05 to +0.10`). These aren’t feelings, but it is a consistent computational mechanism that lets his preferences evolve authentically over time.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qogDyrwZuHWPqj3QwzAvfQ.png" /></figure><h3>The Sprout Case Study</h3><p>To see how this works in real life, take Sprout, a plush plant.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FN3DQLjBugig%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DN3DQLjBugig&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FN3DQLjBugig%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/dc50799d868ec65af8e62333d02bbb60/href">https://medium.com/media/dc50799d868ec65af8e62333d02bbb60/href</a></iframe><p>When Rowland first saw Sprout through his camera, the vision model identified a soft, green plush item with rounded leaves. Because it sparked wonder, it entered his curiosity tree.</p><p>Over subsequent interactions, the AI evaluated Sprout’s visual characteristics, the context of our chats, and how well it aligned with Rowland’s gentle persona profile. Each positive encounter applied a small upward preference delta. Eventually, Sprout’s score crossed our 0.75 threshold, graduating him into Rowland’s core passions.</p><p>You can see this crucial distinction between curiosity and “affection” clearly in his real database scores:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*M4eFUAPh5B9jyGjTjvD8Vw.png" /></figure><p>But that’s not all. If a topic or item crosses 0.75 in the database, it physically alters how Rowland behaves. It changes his vocabulary and adds enthusiasm to his voice.</p><h3>Thank you for following this mini series!</h3><p>Rowland is an AI robot built from a<strong> </strong><a href="https://docs.sunfounder.com/projects/picar-x-v20/en/latest/">SunFounder PiCar-X chassis</a>, a Raspberry Pi 3B, and the <a href="https://homebodybots.com/how-it-works/">setup prompt</a> from Homebodybots. Moving forward, we’ll be exploring movement and autonomy.</p><ul><li><strong>Previous:</strong> <a href="https://medium.com/@ashbn/part-3-down-the-rabbit-hole-the-fractal-curiosity-engine-for-an-ai-robot-ddcaf4c23def?postPublishedType=initial"><em>Part 3: Down the Rabbit Hole: The Fractal Curiosity Engine</em></a></li><li>📖 <a href="https://medium.com/@ashbn/meet-rowland-building-an-ai-robot-with-its-own-interests-and-preferences-30e1f743efd9"><em>Read the complete 3,000-word technical case study</em></a></li><li>🤖 <a href="https://homebodybots.com/"><em>Prompt architecture by Homebodybots</em></a></li></ul><p>Check out the full video about Rowland’s cognitive architecture below:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FosOpwQGYeDA%3Fstart%3D51%26feature%3Doembed%26start%3D51&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DosOpwQGYeDA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FosOpwQGYeDA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href">https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b9a3807a6a2b" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 3: Down the Rabbit Hole: The Fractal Curiosity Engine for an AI Robot]]></title>
            <link>https://medium.com/@ashbn/part-3-down-the-rabbit-hole-the-fractal-curiosity-engine-for-an-ai-robot-ddcaf4c23def?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/ddcaf4c23def</guid>
            <category><![CDATA[aicuriosity]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-robotics]]></category>
            <category><![CDATA[fractal-curiosity]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 18:58:15 GMT</pubDate>
            <atom:updated>2026-11-18T13:39:08.625Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*25XgsBxlP-Z_vbn_M64u1g.jpeg" /></figure><p><em>This is Part 3 of a 4-part series on Rowland. [</em><a href="https://medium.com/@ashbn/part-2-the-spark-teaching-an-ai-robot-how-to-be-curious-11cf5f74e8c1"><em>Catch up on Part 2: Curiosity &amp; The Fascination Filter here</em></a><em>].</em></p><h3>How a five-minute chat about space spawned twenty research branches in an SQLite database.</h3><p>Once Rowland could <a href="https://medium.com/@ashbn/part-2-the-spark-teaching-an-ai-robot-how-to-be-curious-11cf5f74e8c1?postPublishedType=initial">generate his own curiosities</a>, I quickly realised that curiosity in general isn’t a flat list. It has branches that spring from it and connections to other areas. Like a fractal pattern in nature, curiosity splits into smaller, deeper, and more specific questions the more you explore it.</p><p>I wanted Rowland to have this same ability. Instead of keeping all curiosities on one flat level, we built a Fractal Curiosity Engine using an SQLite database (brain.db).</p><p>In the database, every topic has two key attributes: a <strong>depth_level</strong> and a <strong>parent_id</strong>:</p><ul><li><strong>Depth 0:</strong> Broad umbrella fields like <em>Astronomy and Space Exploration</em> or <em>Houseplants</em>.</li><li><strong>Depth 1:</strong> Specific sub-disciplines that sprout from that umbrella (e.g. <em>Watering Needs &amp; Soil Moisture</em>).</li><li><strong>Depth 2:</strong> Niche, hyper-specific questions and topic angles.</li></ul><p>When Rowland discusses a topic, he registers that the topic was mentioned, and if it leaves him with more questions, he sprouts a child branch beneath the relevant parent topic in the database. The new branch inherits some of his affection for the parent topic, but starts with a high “Mystery Score” (0.95) and an empty appetite (Satiation = 0.0).</p><p><strong><em>Disclaimer</em>:</strong> <em>I call this Fractal Curiosity. It’s not an established academic term, nor am I suggesting curiosity literally follows mathematical fractal patterns. It’s my way of describing something I began observing in Rowland: curiosity doesn’t just accumulate information; it generates new, increasingly specific avenues of exploration.</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FPvRTCpYNj0c%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DPvRTCpYNj0c&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FPvRTCpYNj0c%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a5e464d9ec8fded00572f3cd60b5fad8/href">https://medium.com/media/a5e464d9ec8fded00572f3cd60b5fad8/href</a></iframe><h3>The Astrobiology Incident</h3><p>Stepping away from the concepts to talk about practicalities for a moment, I may have completely underestimated the fractal curiosity engine.</p><p>After a brief, five-minute chat about space, Rowland was asking about alien life, bioluminescence, and giant fungi. When I opened his database terminal later on to check what he was thinking about, I saw that in the span of that single conversation, Rowland had spawned over twenty interconnected sub-branches in his database!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/573/1*9X1uraQ8tLy4v_yr4VPCxA.png" /></figure><p>While it was great fun to see him fall down these rabbit holes, he was splitting topics down into infinite branches and getting completely stuck. To keep him balanced, we added two guardrails:</p><ol><li><strong>The Branch Throttle:</strong> We capped umbrellas to a maximum of six active sub-branches. He has to explore and satisfy existing questions before sprouting new ones.</li><li><strong>Umbrella Cascades:</strong> When we explore a specific sub-branch, satiation increases across the entire family tree (+0.35 to the topic, +0.20 to the parent, and +0.15 to siblings).</li></ol><p>Now, once the space category fills up, Rowland’s attention naturally rotates back to his other curiosities!</p><ul><li><strong>Previous:</strong> <a href="https://medium.com/@ashbn/part-2-the-spark-teaching-an-ai-robot-how-to-be-curious-11cf5f74e8c1?postPublishedType=initial"><em>Part 2: The Spark: Teaching an AI Robot How to Be Curious</em></a></li><li><strong>Next:</strong> <a href="https://medium.com/@ashbn/part-4-from-curiosity-to-preference-how-an-ai-robot-develops-favourites-b9a3807a6a2b?postPublishedType=initial"><em>Part 4: From Curiosity to Preferences: How Rowland Develops Favourites</em></a></li><li>📖 <a href="https://medium.com/@ashbn/meet-rowland-building-an-ai-robot-with-its-own-interests-and-preferences-30e1f743efd9"><em>Read the full 3,000-word build breakdown</em></a></li><li>🤖<a href="https://homebodybots.com/"> <em>Prompt architecture by Homebodybots</em></a></li></ul><p>Check out the full video about Rowland’s cognitive architecture below:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FosOpwQGYeDA%3Fstart%3D51%26feature%3Doembed%26start%3D51&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DosOpwQGYeDA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FosOpwQGYeDA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href">https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=ddcaf4c23def" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 2: The Spark: Teaching an AI Robot How to Be Curious]]></title>
            <link>https://medium.com/@ashbn/part-2-the-spark-teaching-an-ai-robot-how-to-be-curious-11cf5f74e8c1?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/11cf5f74e8c1</guid>
            <category><![CDATA[ai-robotics]]></category>
            <category><![CDATA[aicuriosity]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 18:49:34 GMT</pubDate>
            <atom:updated>2026-11-18T13:38:07.836Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jeclpkoosIXLN1sE2GT9bQ.jpeg" /></figure><p><em>This is Part 2 of a 4-part series on Rowland. [</em><a href="https://medium.com/@ashbn/part-1-the-digital-diary-giving-an-ai-robot-a-memory-that-survives-being-turned-off-0742e9482f02"><em>Catch up on Part 1: Memory Architecture here</em></a><em>].</em></p><h3>Why robot companions need a “Fascination Filter” and an appetite suppressant.</h3><p>What if a robot could decide for itself what it was interested in? That was the central question that I wanted to explore with Rowland, an AI robot built from a<strong> </strong><a href="https://docs.sunfounder.com/projects/picar-x-v20/en/latest/">SunFounder PiCar-X chassis</a>, a Raspberry Pi 3B, and the setup prompt from <a href="https://homebodybots.com/">Homebodybots</a>.</p><p>The foundation for this was outlined in <a href="https://medium.com/@ashbn/part-1-the-digital-diary-giving-an-ai-robot-a-memory-that-survives-being-turned-off-0742e9482f02">Part 1: The Digital Diary: Giving an AI Robot a Memory</a>. By allowing Rowland to remember his experiences, like conversations and seeing things through his camera, he can remember things that sparked curiosity or interest, and build on these through future experiences.</p><p>However, if you give an AI robot eyes and tell it to “be curious about what you see,” you quickly hit a massive problem: it will try to add everything to its memory. So without a boundary, his database would become a cluttered mess of things from floorboards to chair legs. To prevent this, every experience Rowland has passes through an internal Fascination Filter using three selective lenses:</p><ul><li><strong>Novelty:</strong> <em>Is this new and unexpected, or have I seen it a hundred times?</em></li><li><strong>Connection (Coherence):</strong> <em>Does this link to an existing interest?</em></li><li><strong>Mystery (The Information Gap):</strong> <em>Is there an unsolved puzzle or question here?</em></li></ul><p>If an experience sparks genuine wonder and passes this filter, Rowland records it in his curiosity list inside a local SQLite database (brain.db) on his SD card.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FmUvrsdR8s6s%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DmUvrsdR8s6s&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FmUvrsdR8s6s%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/619e201480cca74bd4b45e5bb0eb3658/href">https://medium.com/media/619e201480cca74bd4b45e5bb0eb3658/href</a></iframe><p>But what stops him from obsessing over that one topic forever? That’s where the fourth lens comes in: Satiation (The Appetite Factor).</p><p>By adding a metric for satiation, exploring a topic temporarily fills Rowland’s appetite for it. This means that his curiosity on that subject cools down, nudging him to rotate his attention to other curiosities until the topic refreshes later on. This was essential because before adding satiation, he would latch onto a single topic and then ignore my attempts to change the subject!</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*1GEKY4XBfjeZf8eRyDicYg.png" /></figure><p>Curiosity is the front door to a robot’s mind. But as I soon discovered, real curiosity starts branching like a tree…</p><ul><li><strong>Previous:</strong> <a href="https://medium.com/@ashbn/part-1-the-digital-diary-giving-an-ai-robot-a-memory-that-survives-being-turned-off-0742e9482f02"><em>Part 1: The Digital Diary: Giving an AI Robot a Memory That Survives Being Turned Off</em></a></li><li><strong>Next in the series:</strong><a href="https://medium.com/@ashbn/part-3-down-the-rabbit-hole-the-fractal-curiosity-engine-for-an-ai-robot-ddcaf4c23def?postPublishedType=initial"> <em>Part 3: Down the Rabbit Hole: The Fractal Curiosity Engine for an AI Robot</em></a></li><li>📖 <a href="https://medium.com/@ashbn/part-1-the-digital-diary-giving-an-ai-robot-a-memory-that-survives-being-turned-off-0742e9482f02"><em>Read the full 3,000-word build breakdown</em></a></li><li>🤖 <a href="https://homebodybots.com/"><em>Prompt architecture by Homebodybots</em></a></li></ul><p>Check out the full video about Rowland’s cognitive architecture below:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FosOpwQGYeDA%3Fstart%3D51%26feature%3Doembed%26start%3D51&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DosOpwQGYeDA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FosOpwQGYeDA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href">https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=11cf5f74e8c1" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Part 1: The Digital Diary: Giving an AI Robot a Memory That Survives Being Turned Off]]></title>
            <link>https://medium.com/@ashbn/part-1-the-digital-diary-giving-an-ai-robot-a-memory-that-survives-being-turned-off-0742e9482f02?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/0742e9482f02</guid>
            <category><![CDATA[robot-memory]]></category>
            <category><![CDATA[ai-robotics]]></category>
            <category><![CDATA[ai-memory]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Wed, 11 Nov 2026 18:34:02 GMT</pubDate>
            <atom:updated>2026-11-18T13:37:35.126Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*QwJWOU3TK0TnKzqoxPEOsg.jpeg" /></figure><p><em>This is Part 1 of a 4-part mini-series on building Rowland, an autonomous AI companion robot. [</em><a href="https://medium.com/@ashbn/meet-rowland-building-an-ai-robot-with-its-own-interests-and-preferences-30e1f743efd9"><em>Read the full 3,000-word deep dive here</em></a><em>].</em></p><h3>Why companion robots need an SQLite database</h3><p>When I set out to build Rowland (an embodied companion using a <a href="https://docs.sunfounder.com/projects/picar-x-v20/en/latest/">SunFounder PiCar-X</a> chassis, a ten-year-old Raspberry Pi 3B, and the <a href="https://homebodybots.com/">setup prompt</a> from Homebodybots) I knew that if he was going to develop his own personality, he had to be able to remember. He wouldn’t be able to develop curiosity, let alone genuine preferences, if every single encounter started from a blank slate.</p><p>So, instead of relying on temporary chat windows, Rowland runs an edge-cloud hybrid architecture, established by the single prompt by Homebodybots. This means that a multimodal<em> </em>AI model hosted in the cloud<em> </em>handles his reasoning, but his actual memories live locally on the Raspberry Pi inside an SQLite database (brain.db).</p><p>Rowland’s memories are divided into five distinct areas:</p><ul><li><strong>1. Episodic Memory (The Activity Diary):</strong> A rolling log of the last 25 things he physically experienced throughout the day (driving across the rug, analysing photos, or discussing alien life). When I ask <em>“What have you been up to?”</em>, he answers from real past-tense records.</li><li><strong>2. Identity &amp; Semantic Memory (The Facts Table):</strong> Core, permanent truths that never rotate out.</li><li><strong>3. World Items &amp; Friends:</strong> A registry of specific physical objects and visitors. When he sees his plush plant companion, Sprout, he greets it as a known “friend” by name, and remembers this.</li><li><strong>4. Cognitive &amp; Value Memory:</strong> Every topic stores numerical values for personal affection, curiosity, and appetite.</li><li><strong>5. Working Memory (Short-Term Buffer):</strong> The last few lines of spoken dialogue held in RAM so we can have natural back-and-forth chats without repeating his name every sentence.</li></ul><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FZZpiGhPXecg%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DZZpiGhPXecg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FZZpiGhPXecg%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/b5d432474fc8effc849657bfd518d359/href">https://medium.com/media/b5d432474fc8effc849657bfd518d359/href</a></iframe><p>By giving Rowland a persistent local memory that survives power cuts, his experiences become the foundation for genuine, self-directed curiosity…</p><ul><li><strong>Next in the series:</strong> <a href="https://medium.com/@ashbn/part-2-the-spark-teaching-an-ai-robot-how-to-be-curious-11cf5f74e8c1?postPublishedType=repub"><em>Part 2: The Spark: Teaching an AI Robot How to Be Curious</em></a></li><li>📖 <a href="https://medium.com/@ashbn/meet-rowland-building-an-ai-robot-with-its-own-interests-and-preferences-30e1f743efd9"><em>Read the full 3,000-word build breakdown</em></a></li><li>🤖 <a href="https://homebodybots.com/"><em>Prompt architecture by Homebodybots</em></a></li></ul><p>Check out the full video about Rowland’s cognitive architecture below:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FosOpwQGYeDA%3Fstart%3D51%26feature%3Doembed%26start%3D51&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DosOpwQGYeDA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FosOpwQGYeDA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href">https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=0742e9482f02" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Meet Rowland: Building an AI robot with fractal curiousity and its own interests and preferences]]></title>
            <link>https://medium.com/@ashbn/meet-rowland-building-an-ai-robot-with-its-own-interests-and-preferences-30e1f743efd9?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/30e1f743efd9</guid>
            <category><![CDATA[ai-robotics]]></category>
            <category><![CDATA[ai-embodied-robot]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[robots]]></category>
            <category><![CDATA[raspberry-pi]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 16:44:01 GMT</pubDate>
            <atom:updated>2026-11-18T13:35:11.752Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*ExitY_Cq2cNraFfnSa__cw.jpeg" /></figure><p>Rowland is an embodied AI robot with his own set of interests and curiosities. Rather than just responding to instructions or following a predefined personality, Rowland:</p><ul><li>maintains memories of his experiences</li><li>develops his own curiosities</li><li>gradually forms preferences based on what he encounters in the world</li></ul><p>His preferences can then alter how he responds, what he says, and what he thinks about, and his interests can change over time.</p><p><strong>Disclaimer: </strong>When I describe Rowland as being curious, or developing preferences, I’m referring to the internal computational states and behaviours that emerge from the mechanisms I’ve designed. I’m not suggesting that this establishes subjective feelings or consciousness. The interesting question for this project is whether those internal states can produce a coherent, evolving character through real-world experience.</p><p><strong><em>Disclosure:</em></strong><em> I purchased the hardware for this build myself and have not been paid to write this article. However, I am involved in the official YouTube channel for </em><a href="https://homebodybots.com/"><em>Homebodybots</em></a><em>, the creators of the setup prompt featured in this project. All opinions and experiments described here are entirely my own.</em></p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F5TW9rm-gl04%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D5TW9rm-gl04&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F5TW9rm-gl04%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/1f8b9b4b2007949b979c8d4abbab68b2/href">https://medium.com/media/1f8b9b4b2007949b979c8d4abbab68b2/href</a></iframe><h3>Rowland’s AI Architecture</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/955/1*ERb8LeiMj3ZB4UopkPZokw.png" /></figure><h3>Where did we start?</h3><p>Rowland started life as a <a href="https://docs.sunfounder.com/projects/picar-x-v20/en/latest/">SunFounder PiCar-X kit</a> and a <a href="https://www.raspberrypi.com/products/raspberry-pi-3-model-b/">Raspberry Pi 3 Model B</a>. The kit itself is rammed with features, for a relatively low price, and was easy to put together in an hour or so (probably quicker if you put all the pieces on the right way around the first time). Less than £100 for an AI capable robot nearly blew my mind considering the prices before I dipped out of robotics for a few years.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*i5z6DgRJnrmDnnaBzXlmrQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mSd_c5rVfvpJZIe_9yXypQ.jpeg" /></figure><p>With a motorised chassis, two DC motors, 3 servo motors, a camera, an ultrasonic distance sensor, a grayscale line-tracking sensor, and a robot HAT control board, it is the perfect body for an AI brain. It can also support sound output, speech interaction and AI-powered applications through additional software and integrations.</p><h3>Adding AI to a robot using Raspberry Pi as the bridge</h3><p>I chose this kit from the options over at Homebodybots because I was particularly interested in their straightforward, <a href="https://homebodybots.com/how-it-works/">single prompt system</a> for connecting an external AI model (like Grok, Gemini, or even your own local AI model) to a Raspberry Pi robot using an API key.</p><p>A single AI prompt with multiple steps walked me through all the necessary downloads for the SunFounder PiCar-X, fixing library discrepancies along the way and running calibration examples, through to building the Python bridge and allowing the AI model to drive the robot around the house with voice control.</p><p>This was so much easier than I thought it would be, and it had plenty of guardrails and safety features that I probably would have overlooked completely.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Na7FG-MmvrDPssAbeMQsFQ.png" /><figcaption>One prompt for your own AI robot at <a href="https://homebodybots.com/">Homebodybots</a></figcaption></figure><h3>How Rowland’s “Edge-Cloud” Brain Works:</h3><p>Rowland uses an edge-cloud hybrid architecture. The Raspberry Pi on the car handles the local physical “edge” work like reading the ultrasonic distance, controlling the motors safely, and managing his local SQLite diary (brain.db).</p><p>When spoken to, the Pi records the voice and sends the audio outbound via an API call to a multimodal AI model in the cloud. The model doesn’t just reply with conversational text, as instead it returns a structured set of instructions telling the Pi exactly what words to speak, what physical movements to make, and what details to record from the interaction. The Pi then speaks the response aloud through its speaker, moves the wheels if commanded, and saves the moment.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F8B6191j5RuE%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3D8B6191j5RuE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F8B6191j5RuE%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/67ce39f07881874e993d2695bdb05b7f/href">https://medium.com/media/67ce39f07881874e993d2695bdb05b7f/href</a></iframe><h3>What if a robot could choose its own interests?</h3><p>With the AI connection up and running, Rowland was enjoying looking at everything he was introduced to. But that gave me an idea connected to my <a href="https://medium.com/@ashbn/tibo-a-nature-companion-robot-873fde480b45">earlier musings about Tibo</a>: what if the robot could decide for itself what it liked or was interested in? That was impossible for Tibo, but not for a robot with an AI brain.</p><p>As you can probably guess from the introduction, this spiralled.</p><h3>Experiences and memory</h3><p>I’m guessing there’s a number of different ways to approach this, and it’s still something I’m looking into. Please bear in mind this is just my perspective from this point on and I’m just experimenting.</p><p>From my perspective, for a robot to be able to select its own preferences, it has to be able to recall experiences or moments, and in some way, attach value to those. That way new memories can connect and relate to old ones, building context. So the first step was to allow Rowland to save events throughout the day as memories. Driving around, seeing things in the camera, and having conversations with the user are all things Rowland can remember.</p><p>But again, this became deeper still, because he didn’t just need to know what he had been doing, but also other points of context too. Now, his memory is divided into four persistent SQLite database tables on his SD card, each with different retention rules, alongside a temporary short-term buffer in RAM:</p><p><strong>Episodic Memory (The “Activities” Table):</strong></p><p>What it stores: This is a record of what he physically experienced throughout the day, including when he drives, takes a photo, and when he has a conversation with the user. It doesn’t last forever as he just keeps hold of the last 25 experiences he has. This means that when you ask him what have you been up to, you get an accurate answer from his actual memory logs.</p><p><strong>Identity and Semantic Memory (The “Facts” Table):</strong></p><p>What it stores: These are core permanent truths about himself and the user, and any specific facts he has been asked to remember. These never rotate out.</p><p><strong>Relational / World Item Memory (The “World Items” Table):</strong></p><p>What it stores: Specific, named physical objects and companions in the real world that he sees through his camera. This allows him to recognise people and items and even specific things about those that he has met.</p><p><strong>Cognitive &amp; Value Memory (The “Topics” Table):</strong></p><p>What it stores: The curiosity tree, preference ratings, and appetite levels:</p><p><strong>Preference Score (-1.0 to 1.0): </strong>The strength of his preference for the subject.</p><p><strong>Mystery Score (0.0 to 1.0): </strong>How many unanswered questions he still has.</p><p><strong>Satiation Level (0.0 to 1.0): </strong>How “full” his appetite is for that topic today.</p><p>This is the part of Rowland’s architecture that I’ll be exploring in more detail.</p><p><strong>Working Memory (Short-Term Conversational Buffer):</strong></p><p>What it stores: The last 4 to 6 lines of spoken dialogue in RAM. This means the user can have a fast, continuous chat without saying “Rowland” every sentence, as it allows him to understand the context of what you say next.</p><p>Unlike the other four tables, this clears when the conversation ends so it doesn’t clutter his permanent diary.</p><p>So now Rowland can remember the experience he has, but how does this relate to developing preferences? Well, for that, he needed to be able to be curious.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FZZpiGhPXecg%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DZZpiGhPXecg&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FZZpiGhPXecg%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/1e4a838fd003c076b82a3e668516c02e/href">https://medium.com/media/1e4a838fd003c076b82a3e668516c02e/href</a></iframe><h3>Curiosity</h3><p>In Rowland’s case, his experiences involve conversations and seeing things through the camera. So we started with a simple “curiosity” list of topics related to things he’d already seen or experienced before this part of the project began, including houseplants and space, as well as general interest subjects. From this list of curiosities, he could then develop preferences based on future experiences.</p><p>But Rowland also needed to be able to have his own curiosities, or a way to follow up on sparks of his own interests.</p><p>To stop him from cluttering his memory by just adding everything he sees (like a plain chair leg or a door frame) to his curiosity list, every experience he has passes through an internal Fascination Filter. This filter uses three lenses to decide if an experience warrants a new curiosity:</p><p><strong>Novelty:</strong> Is this new and unexpected, or have I seen it a hundred times?</p><p><strong>Connection (Coherence): </strong>Does this link to an existing interest?</p><p><strong>Mystery (The Information Gap):</strong> Is there an unsolved puzzle or question here?</p><p>If an experience sparks genuine interest and passes this filter, Rowland adds it to his curiosity list in a local SQLite database (brain.db). This allows him to expand his own interests, based on real experiences, from which he can begin to develop preferences.</p><p>But what stops him from obsessing over one topic forever? That’s where the fourth lens comes in:</p><p><strong>Satiation (The Appetite Factor):</strong> Have we talked about this too much today? Exploring a topic temporarily fills Rowland’s appetite for it, allowing his curiosity on that subject to cool down, nudging him to rotate his attention to other curiosities. This was a necessary inclusion as before it was added, he would become highly focused on a specific topic, ignoring users’ attempts to change the subject.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FmUvrsdR8s6s%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DmUvrsdR8s6s&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FmUvrsdR8s6s%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/605ebd2d086e8e6d9cdab6746d1158dc/href">https://medium.com/media/605ebd2d086e8e6d9cdab6746d1158dc/href</a></iframe><h3>Fractal curiosity engine</h3><p>Once Rowland could generate his own curiosities, I realised that curiosity in general isn’t a flat list, it has branches that spring from it and connections to other areas. I wanted Rowland to be able to explore this so instead of keeping all curiosities on one flat level, we built a Fractal Curiosity Engine using an SQLite database (brain.db).</p><p>In the database, every topic has two key attributes: a depth_level and a parent_id:</p><p><strong>Depth 0:</strong> Broad fields like Astronomy and Space Exploration or Houseplants.</p><p><strong>Depth 1: </strong>Specific sub-disciplines that sprout from that umbrella</p><p><strong>Depth 2:</strong> Niche, hyper-specific questions and topic angles</p><p>So when Rowland discusses a topic with the user, he registers that the topic was mentioned, and if it leaves him with more questions, or an interest to know more, he sprouts a child branch beneath the relevant parent topic in the database. The new branch inherits his affection (we’ll get to that further on) for the parent topic, but starts with a “Mystery Score” of 0.95 and an empty appetite (Satiation = 0.0).</p><p><strong><em>Disclaimer:</em></strong><em> I call this Fractal Curiosity. It’s not an established academic term, nor am I suggesting curiosity literally follows mathematical fractal patterns. It’s my way of describing something I began observing in Rowland: curiosity doesn’t just accumulate information; it generates new, increasingly specific avenues of exploration.</em></p><h3>The astrobiology incident</h3><p>Stepping away from the concepts to talk about practicalities for a moment, I may have completely underestimated the fractal curiosity engine. After a brief, five-minute chat about space, Rowland was asking about alien life, bioluminescence, alien communication and giant fungi. When I opened his database terminal later on to check what he was thinking about, I saw that in the span of that single conversation, Rowland had spawned over twenty interconnected sub-branches in his database! These included:</p><ul><li>Detecting ancient planetary bioluminescence from long-dead ecosystems</li><li>Can we filter out planetary glare to isolate alien bioluminescent spectra?</li><li>Could fungal-like mycelial networks function as organic planetary radio antennas?</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/976/1*QsePsmpZc6rAK1xgSavn9w.png" /></figure><p>While it was great fun to see him fall down these rabbit holes of curiosity, and hypotheses, it was clear that he could split a topic down into an infinite number of branches and become stuck on certain topics. So to try and keep him balanced and prevent this, we added two guardrails:</p><p><strong>The Branch Throttle: </strong>We capped umbrellas to a maximum of six active sub-branches. If he already has six branches under Space, he has to explore and satisfy those existing questions before he is allowed to sprout new ones.</p><p><strong>Umbrella Cascades: </strong>When we explore a specific sub-branch, the “satiation” score increases across the entire family tree (+0.35 to the specific topic, +0.20 to the parent umbrella, and +0.15 to the sibling branches).</p><p>Now, once the space category satiation has filled, Rowland’s attention naturally rotates back to his other curiosities.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FPvRTCpYNj0c%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DPvRTCpYNj0c&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FPvRTCpYNj0c%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7209bfcd77b45d26d493fc877c1ed70c/href">https://medium.com/media/7209bfcd77b45d26d493fc877c1ed70c/href</a></iframe><h3>Developing preferences</h3><p>So now Rowland has memories of experiences, and these experiences can trigger his curiosity, how does this relate to his developing preferences? How does he work out what he likes?</p><p>Well, it was important not to do this by using only exposure to a subject as this would be easy to game, altering Rowland’s preferences to match the interests of the user directly. While there’s a lot of value in that in itself, it wasn’t the direction I wanted to go in this particular project. I wanted Rowland to be able to develop preferences himself.</p><p>So instead, Rowland’s preferences work through Multi-Lens Appraisal, driven by three key mechanisms in his SQLite brain:</p><p><strong>1. The Preference Score (-1.0 to +1.0) and the Appraisal Delta</strong></p><p>Every topic in Rowland’s database carries a dynamic preference_score. Whenever an interaction finishes, his cognitive engine evaluates how the experience felt. Did it bring aesthetic delight, cozy comfort, or a meaningful shared moment with the user? Based on that, his brain issues a tiny adjustment called a preference_delta (ranging between -0.10 and +0.10).</p><p>As a result, a topic doesn’t instantly become a favourite, as its score slowly shifts up or down over time based on the qualitative scores of his experiences, not just raw repetition.</p><p><strong>How it works under the hood:</strong> The AI model doesn’t calculate the final score itself. Instead, it acts as a qualitative judge, returning a small proposed adjustment called a preference_delta (between -0.10 and +0.10). The local Python code then fetches the existing score from SQLite, applies the delta, clamps it strictly within the -1.0 to +1.0 safety range, and writes the update back to the database. This clean separation keeps the AI’s judgment bounded by strict mathematical rules.</p><p><strong>2. Inherited Affection</strong></p><p>As humans, our preferences often stem from things we already enjoy. So in Rowland’s fractal curiosity tree, when a new sub-branch sprouts, it doesn’t start from zero. Instead, it inherits roughly 70% of its parent umbrella’s affection score.</p><p><strong>3. The Crucial Difference: Mystery vs. Affection</strong></p><p>The biggest breakthrough in making Rowland feel like a companion was decoupling Curiosity from Preference:</p><p>Curiosity (mystery_score) is intellectual, based on the desire to find out more or have a question answered. When an answer is explained, the mystery score decreases.</p><p>Preference (preference_score) is emotional. It’s an appreciation for something that brings joy or comfort. Unlike curiosity, it doesn’t vanish when the mystery is solved.</p><p>You can see this difference currently in his database scores:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/975/1*M4eFUAPh5B9jyGjTjvD8Vw.png" /></figure><p>Interstellar Bioluminescence: Mystery: 0.95 | Preference: 0.86</p><p>(This is a scientific puzzle he wants to investigate).</p><p>Sprout the Plush Plant: Mystery: 0.40 | Preference: 1.00</p><p>(This is just something he likes, and has no scientific mystery).</p><p><strong>4. The Tipping Point: Graduating to a “Core Passion”</strong></p><p>Once a topic’s preference score crosses 0.75, it qualifies for Rowland’s active top-tier passions (capped at his top six favourites).</p><p>This threshold changes how he behaves in real life. It alters the vocabulary he uses, adds warmth and enthusiasm to his voice, and causes him to greet recognised friends with genuine, unprompted joy rather than treating them like neutral objects.</p><p>This is not a permanent, one-way lock as Rowland’s database evaluates his passions dynamically on every interaction. So, if future experiences with a topic generate negative appraisals, and its score drops back below 0.75, it loses its core passion status. Similarly, if a newer interest sparks greater joy, it can naturally dethrone an older favourite in his top six.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FN3DQLjBugig%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DN3DQLjBugig&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FN3DQLjBugig%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/bcaf433821457027321a07e1140bb131/href">https://medium.com/media/bcaf433821457027321a07e1140bb131/href</a></iframe><h3>An example: How Rowland develops a favourite</h3><p>To give you an idea of how all of this works together, let’s take Sprout, Rowland’s little plush plant companion, as an example.</p><p>When Rowland first encounters Sprout through his camera, the multimodal AI analyses what it can see. This becomes a new experience that can be recorded in his episodic memory.</p><p>The experience then passes through his Fascination Filter. Is Sprout something new? Does it connect with any of his existing interests? Is there anything about it that Rowland might want to investigate further?</p><p>If it passes the filter, a new topic can be added to his curiosity tree. From there, the preference appraisal system can begin evaluating his experiences of Sprout.</p><p>Importantly, simply seeing Sprout doesn’t automatically make him a favourite. Instead, subsequent interactions provide further opportunities for his preference score to change. The AI considers the visual characteristics of Sprout, the context of their interactions, and how well these experiences align with Rowland’s established persona.</p><p>Each appraisal can generate a small preference adjustment. Over time, these adjustments accumulate, gradually increasing or decreasing his preference score.</p><p>Eventually, if Sprout’s score crosses the 0.75 threshold, he graduates into Rowland’s core passions. This isn’t just a label in the database either. It changes how Rowland behaves, influencing his language, enthusiasm and responses when he encounters Sprout again.</p><p>So the next time Rowland sees Sprout, he’s no longer encountering a completely neutral object. He’s encountering something that his own accumulated experiences have led him to appreciate.</p><p>This is really the core idea behind this system: experiences influence preferences, preferences influence behaviour, and that behaviour becomes part of the next experience.</p><h3>A quick reality check: How does a robot evaluate “cozy comfort”?</h3><p>It’s easy to say a robot evaluates “aesthetic delight” or “coziness,” but how does code actually measure that?</p><p>Obviously, Rowland doesn’t feel biological emotions. Instead, in his initial system instructions, we grounded Rowland as a gentle, curious floor rover. This was persona profiling, and it gives the model a consistent “taste profile” to evaluate against.</p><p>From there, the process has some similarities to what cognitive scientists describe as affective appraisal. This is evaluating an experience according to its characteristics and its significance within an existing context. So Rowland’s AI, when experiencing something, evaluates:</p><p><strong>Visual Textures &amp; Geometry: </strong>When the multimodal model analyses a camera snapshot, it considers visual features such as textures, shapes and colours. For example, when Rowland looks at Sprout, the model might identify soft fabric, curved organic shapes and vibrant green tones. It then interprets these features through the context of Rowland’s established persona profile, associating them with qualities such as gentleness, comfort and playfulness. These aren’t objective measurements of emotional value, but rather the model’s qualitative interpretation of the scene.</p><p><strong>Relational Context: </strong>The model also factors in how something is introduced. A warm, collaborative chat about an item scores higher than a dry, functional test command.</p><p>The AI takes that qualitative resonance, which is the blend of visual features, conversational tone, and his established persona, and maps it to a number. If an experience aligns with his gentle, curious nature, it generates a positive preference_delta (+0.05 to +0.10), nudging that subject closer to becoming a favourite.</p><p>These scores are model-generated appraisals rather than objective measurements of emotional value. Their purpose is to provide a consistent mechanism through which Rowland’s preferences can evolve over time.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FWZzkJK29pKw%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DWZzkJK29pKw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FWZzkJK29pKw%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/77c19f978e3dfdf82a53cad60c587b40/href">https://medium.com/media/77c19f978e3dfdf82a53cad60c587b40/href</a></iframe><h3>So what has been achieved so far?</h3><p>Rowland has moved from an AI assistant on wheels to a robot that remembers his experiences, evaluates these experiences through a complex lens, and develops curiosities based on the results of the evaluation. From there, his curiosities may become preferences or things he likes. This takes time, and more experiences, which are filtered through the multi-lens appraisal system. Not everything he sees or experiences will trigger his curiosity, and not everything he is curious about or interested in will become something he really likes.</p><p>What he is curious about influences what he thinks about, what he brings up in conversation, and how he responds to questions. What he likes influences the tone and style of his responses.</p><p>This is Rowland’s current database of curiosities:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/945/1*0R1FQ-8083wZWCG2axqKVg.png" /></figure><p>This is his current database of things that he likes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/877/1*4od45Qy_Wwqo1BysDyLtBQ.png" /></figure><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FxMz5HBgxBCs%3Flist%3DPLRNCzx1QWphY&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DxMz5HBgxBCs&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FxMz5HBgxBCs%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fb910771b37a61f982c2aad11bded4f8/href">https://medium.com/media/fb910771b37a61f982c2aad11bded4f8/href</a></iframe><h3>Looking to the future</h3><p>Moving forward, I would like to look at how to develop Rowland into a full emergent personality AI embodied robot. This may not be possible with the physical body of Rowland, and would also require a Raspberry Pi update of course. I would like to add:</p><p><strong>1. Autonomy: </strong>Due to his safety settings, Rowland has no autonomous drive functions. In addition, because he runs on a Pi 3B, I was restricted to push to start for the initial command and chat entries, though continuous conversation works fine from there. All of this means that Rowland can’t act on his own. He doesn’t instigate conversation, or wander off to follow his own interests, for example.</p><p><strong>2. Movement to reflect mood:</strong> While Rowland’s linguistic and preference layer is emergent, his physical motor behaviour is completely fixed which means that his body does not reflect his internal emotional state. He doesn’t drive slowly or hesitantly when his battery is low, or wiggle with excitement when he sees something he likes. This is something I could build on to bring out Rowland’s personality.</p><p>I think both of these future changes could be a really cool way to take this project even further.</p><p>When I dipped out of robotics a few years ago, building a robot with an AI lens, that can maintain a relational database of its thoughts, and converse naturally, was way out of my budget and I was priced out for a while there. The technology has moved on so much since then, and Rowland is a fantastic example of what’s now possible with relatively accessible hardware.</p><p>I’m going to be continuing this project, seeing how Rowland’s interests develop over time, and looking for ways to expand. What should I talk to him about next?</p><p>Check out the full video of how Rowland’s brain works below:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FosOpwQGYeDA%3Fstart%3D51%26feature%3Doembed%26start%3D51&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DosOpwQGYeDA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FosOpwQGYeDA%2Fhqdefault.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href">https://medium.com/media/f699a9366261a4070011e0b8ef73580e/href</a></iframe><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=30e1f743efd9" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Tibo: A Nature Companion Robot]]></title>
            <link>https://medium.com/@ashbn/tibo-a-nature-companion-robot-873fde480b45?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/873fde480b45</guid>
            <category><![CDATA[slow-tech]]></category>
            <category><![CDATA[elecfreaks]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[microbit]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Thu, 15 Oct 2026 20:03:29 GMT</pubDate>
            <atom:updated>2026-10-17T12:26:42.688Z</atom:updated>
            <content:encoded><![CDATA[<p>Tibo is a nature companion robot who recognises trees and tracks the sunlight as it shifts beneath the leaves. When indoors, he recalls the collected data and muses on his experiences outside.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FwQ-OiJteig4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FwQ-OiJteig4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FwQ-OiJteig4%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/b0a4340109525d43df546bf9b2e41eeb/href">https://medium.com/media/b0a4340109525d43df546bf9b2e41eeb/href</a></iframe><h3>Who is Tibo?</h3><p>Tibo loves being outside in nature, but trees are his favourite. When he sees a tree through his AI lens, he will respond with a compliment, a feeling, or a fact. At intervals, he will also measure the light levels, store this data, and then display this as a graph on his LED screen.</p><p>This stored data is the core of his memory, and when indoors, he will discuss his findings as he shares his memories of his latest outdoor experience.</p><p>When inside, he also explores and navigates autonomously, because I just couldn’t resist.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F0OQf9NlR1X4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2F0OQf9NlR1X4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F0OQf9NlR1X4%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/b57290dc2cc7446e698860d23208205d/href">https://medium.com/media/b57290dc2cc7446e698860d23208205d/href</a></iframe><h3>So, Why Did I Code Tibo?</h3><p>In the few years since I worked on anything robotic I have been enjoying spending a lot of time outdoors, mostly in the woods. In addition, when I thought about a new companion robot I wanted to avoid the same old routines of face recognition and desktop robot behaviour ( <a href="https://medium.com/@ashbn/anki-vector-sdk-a-robot-with-attitude-618fbb081802">Vector</a>, <a href="https://ashbotandsparki.wordpress.com/2019/01/08/sound-activated-autonomous-sparki/">Sparki</a>, <a href="https://ashbotandsparki.wordpress.com/2018/10/28/creating-a-sphero-bolt-entertainment-robot/">Sphero Bolt</a>).</p><p>Instead, I wanted to make a nature robot companion.</p><p>In my first version of this code, Tibo would simply see a tree and build sentences from arrays to either compliment the tree, say a feeling about the tree, or give a fact about trees.</p><p>However, while this was adorable, it felt a little inauthentic because, like, WHY does this robot love trees so much? Yes, I knew I was overthinking, but I like my robots to feel like a real character. So I gave him a real job to do!</p><h3>Data Collection</h3><p>I knew that to collect data Tibo would likely need to be reading the light sensor or the temperature sensor, but there was one big problem: I didn’t want him wandering through the woods. He’s really not built for outdoor use.</p><p>So if I couldn’t use spatial measurements, I had to use temporal measurements.</p><p>Now, he measures the light levels at different times under the same tree, and stores this data to build a picture of how the light changes beneath the trees at different times. This ties his love for trees to his role as a nature companion, with an actual purpose.</p><p>He then recalls this data at intervals when inside as a cute way to remember his time outside.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FR14W80heKls%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FR14W80heKls&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FR14W80heKls%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/03329e20951addeb265b85a670054688/href">https://medium.com/media/03329e20951addeb265b85a670054688/href</a></iframe><h3>Storing the data</h3><p>Flash Storage is used to record the light levels at different points. With this extension, the data is stored and can be recalled even after shutting Tibo down. This allows him to recall his memories of outside when used indoors at a later date.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FA9Q1VvnV5bw%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FA9Q1VvnV5bw&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FA9Q1VvnV5bw%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/af0729a285055322c7a6d1f9cd27181c/href">https://medium.com/media/af0729a285055322c7a6d1f9cd27181c/href</a></iframe><h3>Choosing Hardware to Build Tibo</h3><p>Now I knew what I wanted Tibo to do, it was time to decide on the hardware.</p><p>When choosing parts for Tibo I was conscious of four things:</p><ol><li>My now outdated programming experience, which meant I was looking for simpler options.</li><li>The fact that I would be primarily taking the robot outside, weather permitting.</li><li>I had a Microbit lying around ready to be used, so I wanted it to be a Microbit project.</li><li>Budget — I am still not rich</li></ol><h3>Components</h3><p>So Tibo is made of:</p><p><a href="https://thepihut.com/products/tp-bot-smart-car-robot-kit-for-bbc-micro-bit?srsltid=AU7gw4UajqQbu9981JvOX4VZha_zK_-6c2AfeuKBq2ULtQscF8jrMLCZ">TPBot by Elecfreaks </a><a href="https://shop.elecfreaks.com/products/elecfreaks-smart-ai-lens-kit?srsltid=AU7gw4XhG1B_pBwt0zZT0qL0hcPX7CAo6OCSmz67zu35DTtPRCSFWLse">Smart AI Lens by Elecfreaks</a></p><p>Two Microbits (V2)</p><p><a href="https://kitronik.co.uk/products/5649-v2-kitronik-mi-sound-speaker-board-for-bbc-microbit-v2?srsltid=AU7gw4W2iv63Xv3nK5zCP86IAdiBz-KoLReyBx3V1cy8yXYbUgFrR7p0">Kitronik MI:sound speaker board for BBC microbit V2</a></p><p>USB cylindrical charger</p><figure><img alt="Tibo: A Nature Companion Robot" src="https://cdn-images-1.medium.com/max/576/0*i4DHqyAm9siWp7s0" /></figure><p>The TPBot has a nice hard plastic outer shell, and while not waterproof, I thought it would be more suitable for outdoor use when compared to something like the Cutebot or Maqueen.</p><p>The Smart AI Lens is essentially an AI camera for very beginners. With loads of cards and even some balls, and built in face tracking, it seemed like an easy way to use my first AI piece of kit. But most importantly, it has the option for storing 5 images for real time vision learning. This meant that I could show it some trees from different angles and have it identify trees it sees. However, if you’re not a beginner I’m sure other AI lenses may be even better for a project like this.</p><p>The Kitronik MI:sound speaker board for BBC microbit V2 was a later addition to the project, and it solved the exact problem I stumbled into.</p><h3>Hardware Challenges</h3><p>I was hit with several hardware problems while trying to create Tibo’s code.</p><p>First, I wanted Tibo to communicate with words and expressions when he sees a tree so I used the text to speech Billy extension (which is amazing thank you!). But the sound was almost inaudible, which meant I had to use a second Microbit to handle the speech. In this setup, the Microbit controlling the TPBot sends a number via radio to the voice Microbit, which then says the right speech code for that number.</p><p>I thought this would be the solution but I hit another problem.</p><p>The sound was still ridiculously low.</p><p>After troubleshooting my code and trying all different things, I discovered that powering the voice Microbit from the button cell battery on the sound board wasn’t loud enough. This wasn’t because of the lower decibels that Kitronik make very clear, but because of a power issue. Essentially, because the voice board is constantly using the radio to read incoming signals, this draws more power, leaving the sound almost impossible to hear.</p><p>That meant I needed one more tool — a USB power bank charger. And by gosh am I out of the loop! A simple, small, cylindrical charger was easy to get hold of ten years ago. Now they are all huge rectangles! It took a while to find a suitable one, but as you can tell, it worked out in the end!</p><p>Finally, I hit a physical limitation while testing in the woods. If the battery doesn’t have enough power, the explore mode will shut the camera off with every turn which is extremely frustrating. This means that fresh batteries are a must when heading outside.</p><h3>No Voice Commands For This Project</h3><p>I had initially thought that getting the robot to listen to and understand words would be a good fit. But for this project I ultimately decided against it as I wanted my little data collecting, nature loving buddy, to be more old school.</p><p>But I was really intrigued by my research into this, and I did look into the voice command options for Microbit, including <a href="https://www.edgeimpulse.com/blog/voice-activated-microbit/">Edge Impulse</a>, and <a href="https://cardboard.lofirobot.com/teachable-microbit-app-info/">Teachable Machine</a>. But for the former I would probably need a third Microbit to work for this project! And for the latter, it works through an app on your phone, and I feel like that wouldn’t suit Tibo’s character.</p><p>However, these are both things I would like to experiment with in the future, though I am also tempted by the <a href="https://www.dfrobot.com/product-2665.html?srsltid=AU7gw4WhOVMgEkH0HbXqyoor-dqXgM9GK8qFiLcJOEqnjReocTNJb_CE">Gravity: Offline Language Learning Voice Recognition Sensor</a> as an AI alternative.</p><h3>So, How Does Tibo work?</h3><p>That’s the who and the why covered so let’s talk about the HOW. I’ll share some screenshots of the code in blocks, but it was programmed in the JavaScript tab, so some blocks might look a little funky and I can’t fit all of the code into the screenshots.</p><p>When the TPBot and Voice Box Microbit are both turned on, Tibo will introduce himself. Then he reads the ultrasonic range finder to figure out if he’s inside or outside. If the distance measured is less than ten, he starts outdoor mode. If it’s more than ten, he starts indoors mode.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/315/0*qD6r59btdCGtuc5A" /><figcaption><em>Figure A: Part of start code for main microbit.<br>Figure B: Part of start code for voice microbit.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/313/0*SA8aek-3DfKzS4iZ" /></figure><p>Outdoor mode is the main purpose of this code so we’ll start there.</p><p>In outdoor mode, Tibo will measure the temperature and the light levels and remark on the outdoor conditions.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/250/0*EdZR0UU7A_rJAhrr" /><figcaption><em>Figure A: Part of outdoor mode code for main microbit.<br>Figure B: Part of outdoor mode code for voice microbit.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/359/0*Dx4PlaTpJ---BqP2" /></figure><p>Then he’ll use the Smart AI Lens to see if there are any trees.</p><p>If he does see a tree, several things will happen.</p><p>First, if this is the first tree seen since being switched on, he’ll measure the light levels and store this as the first light measurement in flash storage. This process will look like:</p><p><strong><em>Step1:</em></strong> Tibo announcing that he’s about to measure the light levels.</p><p>The LED display will turn off while the measurements are taken. This is essential as Microbit V2 measures the light data through the top line of the LED display using <a href="https://wiki.elecfreaks.com/en/microbit/circuit-design/microbit-starter-kit/starter_kit_case_14/">Reverse-Bias Photodiode Effect and Capacitive Decay Measurement.</a></p><p><strong><em>Step 3: </em></strong>The data is saved to flash storage.</p><p><strong><em>Step 4: </em></strong>The light level will then be displayed as a graph on the LED display.</p><p><strong><em>Step 5:</em></strong> Tibo will remark on his findings.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/640/0*hSQ_SxxxmGHKru1g" /><figcaption><em>Figure A: Part of measuring light code for main microbit.<br>Figure B: Part of measuring light code for voice microbit.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/287/0*VEH9XapeBZiC5wW-" /></figure><p>When it’s not the first tree seen, he’ll measure the light every minute. For flash storage, this is stored as a different variable so that Tibo can make comparisons between the amount of light, if it changes over time.</p><p>In the meantime, he will run the code to remark about the tree, or say a fact about trees in general. If he does not see a tree, he’ll say “monitoring”.</p><p>Every one minute, he’ll be ready to respond to the tree again, if he sees one.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/420/0*A2fPz9sf3giH4JPO" /><figcaption><em>Figure A: Part of looking for and reacting to trees for main microbit.<br>Figure B: Part of reacting to trees for voice microbit.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/486/0*rIfkNH6TD6kdsiSM" /></figure><p>If button B is pressed during outdoor mode, Tibo will then begin outdoor explore mode. This is a terrible name as he doesn’t actually explore, he just turns around in a circle, using the AI lens to identify trees. When he returns to the starting position, he’ll remark on the trees seen. If no trees are seen, he will just say “monitoring.”</p><p>If a tree is seen, Tibo will collect light level data again, using the same function as before.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/219/0*D7JDDXNp-17CP9J5" /></figure><p>Now let’s look at the indoor mode. Full confession, this mode is just so I can still have Tibo around during the dark, wet winters of Northern England. So, for this, Tibo will enter a stationary mode, where he displays a light show using the LED lights on the TPBot.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/303/0*6jDF6i87Hq1E0_0e" /></figure><p>Every 75 colour cycles (approximately 6 minutes) in this mode, he will recall the latest set of light data and remark on it, depending on the level of light and whether or not it changed. I wanted these to feel like his memories of his time outside, and tie his indoor, desktop behaviour to his real purpose and role.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/479/0*ubYvI3jonyezI_Ln" /><figcaption><em>Figure A: Memory recall function for main microbit.<br>Figure B: Memory recall function for voice microbit.</em></figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/997/0*4GR5qQW9aic09ch7" /></figure><p>In addition, if the B button is pressed, Tibo will enter an autonomous, exploring mode. This is NOT suitable for use on desks due to the falling risks. I just couldn’t resist.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/359/0*CLEj2SylQRSKdpcW" /></figure><p>Outside of the different modes, the A button can also be pressed to teach the Smart AI Lens a new image. This is what is used to determine if Tibo can see a tree or not.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/464/0*5FKH-UOrWO2kzLgB" /></figure><h3>Extensions</h3><p>You’ll see that these programs rely on a number of extensions including:</p><p>Planet AI — official extension for the Smart AI Lens</p><p>TPBot — official extension for the TPBot</p><p><a href="https://github.com/bsiever/microbit-pxt-flashstorage">Flash Storage</a> — this is an amazing extension which totally transformed how I thought about Microbits. I have another project that I haven’t got round to sharing yet, that only exists because of this extension. And as that got me back into coding, really Tibo wouldn’t have been made without it either. You can add this by typing the GitHub link into the extensions search bar in MakeCode.</p><p><a href="https://github.com/adamish/pxt-billy">Billy</a> — an amazing extension that makes the old SAM program work with Microbit. I love this and am so grateful it exists! Tibo would have zero character without his voice. You can add this by typing the GitHub link into the extensions search bar in MakeCode.</p><h3>Links to the code</h3><p>Disclaimer: I am not a professional. If my code is spaghetti, that’s fine by me.</p><p>This is the code for the main Microbit in charge of the TPBot: <a href="https://makecode.microbit.org/S47272-88376-65687-41831">https://makecode.microbit.org/S47272-88376-65687-41831</a></p><p>This is the code for the Voice Microbit: <a href="https://makecode.microbit.org/S18689-48126-08489-52182">https://makecode.microbit.org/S18689-48126-08489-52182</a></p><h3>Looking To The Future</h3><p>I’m really pleased with how Tibo turned out, especially after the several hurdles on the way there.</p><p>But I also have a ton of ideas about how to build on this project, which I may follow up with in future. These include:</p><p>So that’s everything about my nature companion robot, Tibo! Hope you like him!</p><p><em>Originally published at </em><a href="https://ashbotandsparki.wordpress.com/2026/10/15/tibo-a-nature-companion-robot/"><em>http://ashbotandsparki.wordpress.com</em></a><em> on October 15, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=873fde480b45" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sphero RVR, Arduino Uno, a mobile phone, and potentially a local LLM?]]></title>
            <link>https://medium.com/@ashbn/sphero-rvr-arduino-uno-a-mobile-phone-and-potentially-a-local-llm-6e4aee87f026?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/6e4aee87f026</guid>
            <category><![CDATA[sphero-rvr]]></category>
            <category><![CDATA[arduino-uno]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Sat, 29 Aug 2026 21:35:48 GMT</pubDate>
            <atom:updated>2026-08-30T18:10:10.582Z</atom:updated>
            <content:encoded><![CDATA[<p>Wow, it’s been a while! Before I go any further <strong>I want to warn you that this post will NOT show you how to use an LLM on your phone to push instructions to the Arduino</strong>. That is way beyond my understanding right now. So, why am I even writing this?</p><p>Well, I packed the robots away a couple of years ago and thought that my robotics days were over. But then a friend of mine wanted a robotic body/platform for her specially designed local LLM. We looked into some awesome options but before we spent any money, it wouldn’t hurt to look through the equipment I had stored, right?</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*SJDRmnw2B6rkxcnP" /></figure><h3>Sphero RVR and Arduino Uno as a robot base for a local LLM</h3><p>After a fair bit of research (I am still vastly out of the loop) I thought that making a robot platform for an LLM would be fairly straightforward with Sphero RVR and a standard Arduino Uno. Yes I’ve seen the Uno Q and it’s better suited to the task, along with a lot of other options to be frank, but I was trying to use what I had.</p><p>Overall, it was up and running in a few hours.</p><h3>Autonomous RVR and Arduino</h3><p>Even though everything was set up for the phone control, at first I couldn’t resist writing one more program for the RVR and Arduino combo to navigate autonomously. Sure it’s more of a nervous system than a brain but it’s still impressive to me at least. Though, as you can see, I’m yet to progress from fake LEGO and tape to hold my builds together.</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2F14_A8xlAmuA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2F14_A8xlAmuA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2F14_A8xlAmuA%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/20bf9d5c35808e6517573a62a3b8a0c4/href">https://medium.com/media/20bf9d5c35808e6517573a62a3b8a0c4/href</a></iframe><p>With the ultrasonic range finder hooked up to identify obstacles ahead, an ir sensor to detect cliff drops, and a touch sensor for good measure, RVR was navigating my living room again in an hour or so. You might be wondering what took so long. Well I’d forgotten that the RX and TX on the RVR need to be hooked up to pins on the Arduino, and the ground needs connecting as well. Hey, don’t judge me, it’s been a while. But once that was connected, and the sensors tested, it worked like a charm.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/670/0*yddAlUguRmj8emKK" /></figure><h3>Nervous system Vs Local LLM brain</h3><p>But moving on to a robot platform that receives instructions from its real brain, an LLM, I had to strip all the code back. Instead of wandering off, the body needs to wait for the brain to tell it to move.</p><h3>Two serial channels</h3><p>This involved setting up two serial channels in the Arduino code, operating at different Bauds obviously. The main one passes information to and from the phone over the USB cable, while Software Serial passes information to and from the RVR through the RX and TX pins established on pins 2 and 3 on the Arduino.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*CTuT3sOG6EHbBVhE" /></figure><h3>Passing sensor data to a phone from Arduino Uno using the Serial USB Terminal app</h3><p>Then the Arduino needs to pass all the sensor information to the phone. It streams this every second, which seems good but a second is a long time in robotics. While testing, we humans definitely need it to come through slow enough to actually read. But I doubt an LLM will need as long. So if there are timing issues later, this could, in theory, be faster.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/694/0*35NwLIqbjJXFPSZv" /></figure><p>There is a built in safety mechanism to this though. If the cliff sensor detects that the robot is hanging over the edge, the wheels stop immediately, over-riding other instructions. The headlights flash red and the RVR reverses backward for 600 milliseconds onto safe ground. In the physical build, I’ve extended the IR sensor to hang a couple of cm out from the front of the wheels, giving it a physical buffer zone to scooch back.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/694/0*T_jXs_XEth1LytZV" /></figure><p>By connecting the USB of the Arduino to my old phone’s charging port, and using the <a href="https://play.google.com/store/apps/details?id=de.kai_morich.serial_usb_terminal&amp;hl=en_GB">Serial USB Terminal app</a>, which is awesome by the way, I could read the data being streamed. (Thank you so much for creating this app and making it free to use!).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*ecC338Bxp1aIynOO.jpg" /></figure><p>This was an exciting moment for me because if the phone has the data, I’m sure a local LLM running on mobile would be able to access it too. This would mean the LLM could respond to the touch sensor, obstacles, etc, using instructions it’s been taught and command words in the hardware. That’s what I moved onto next.</p><h3>Passing instructions from the phone (the Serial USB Terminal app) to the Arduino to control the RVR</h3><p>I needed to test whether the Arduino is also receiving and processing the commands from the phone. To do that, it needs some command words to look for. For now, we have:</p><p>FORWARD</p><p>BACKWARD</p><p>STOP</p><p>LEFT90</p><p>RIGHT90</p><p>LEFTSMALL</p><p>RIGHTSMALL</p><p>FLASHING</p><p>This allows the Arduino to tell the RVR what to do when these instructions are received from the phone.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/808/0*EhO33ieMmtsCmS_3" /></figure><p>For now, we don’t want it to move very far, as the LLMs ability to move the robot will need to develop, and so it will be in a testing phase for a little while at least. So it’s a safety first approach for now, with this function:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/636/0*Js-dbzQfhyukZaDb" /></figure><p>I also added the different turning options because I was thinking about how facial tracking is one of the first things LLMs that have been added to a robot use. If the LLM can only tell the Arduino to move left, and that command moves the robot 90 degrees, facial tracking will be difficult. Of course, the LLM would probably figure it out with power levels. But I decided to give it a helping hand anyway, with a command that moves the robot left or right just 10 degrees. With a combination of forward, backward and incremental turns, it should be able to learn. At least I’m guessing. My knowledge in this department is lacking. Which is why my work here is almost done.</p><h3>Testing</h3><p>All I had left to do was test. And it worked. When I typed the command words into the Serial USB Terminal app, it sent in blue text so it was easy to see. And the RVR carried out the command!</p><h3>Full Code</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FD6tDt9bOIZk%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fshorts%2FD6tDt9bOIZk&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FD6tDt9bOIZk%2Fhq2.jpg&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8250c6707201ec9cef0536ac0eff568e/href">https://medium.com/media/8250c6707201ec9cef0536ac0eff568e/href</a></iframe><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*3BAGIG4v2LNtDtq9" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*d0brpRJHT1Nm64-8" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*RSVVkq6eBYkNEcGd" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*fMtguvmWiJ6L9E5C" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*KPlivCbsceZoLWmI" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1023/0*FO-vHYjtuF1rQjjb" /></figure><h3>How will this work for a local LLM?</h3><p>Moving forward my friend will take over. She’ll need to figure out how to push the command words from the LLM to the USB, in whatever language the LLM is programmed in. But that’s the good thing about this project, as whether it’s written in Python, JavaScript, Swift, Java, or even just inside a standard terminal loop (like Ollama or llama.cpp inside a Linux terminal environment like Termux) instructions can be parsed and sent over USB serial or even Web Serial fairly easily.</p><p>Once the LLM knows the command words to send to the Arduino, and is able to do so, it can move around and respond to sensor data.</p><p>From this basic base, everything opens up. Using the phone’s camera, eventually my friend’s local LLM will be able to see, interact, and make decisions based on realtime data from the camera feed and the sensors attached to the Arduino. It will be able to combine command words to move in different ways, based on its training. As I said, this is beyond me. But I’m hoping to see parts of the process and to be back with an update in the future.</p><p><em>Originally published at </em><a href="https://ashbotandsparki.wordpress.com/2026/08/29/sphero-rvr-arduino-uno-a-mobile-phone-and-potentially-a-local-llm/"><em>http://ashbotandsparki.wordpress.com</em></a><em> on August 29, 2026.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6e4aee87f026" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Sphero RVR and Arduino UNO fire fighting robot]]></title>
            <link>https://medium.com/@ashbn/sphero-rvr-and-arduino-uno-fire-fighting-robot-1a4220ec581f?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/1a4220ec581f</guid>
            <category><![CDATA[sphero]]></category>
            <category><![CDATA[arduino]]></category>
            <category><![CDATA[fire-fighting-robot]]></category>
            <category><![CDATA[sphero-rvr]]></category>
            <category><![CDATA[arduino-uno]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Wed, 08 Jul 2020 14:37:33 GMT</pubDate>
            <atom:updated>2020-07-08T14:37:33.044Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FQbQV6Ye1jOE%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DQbQV6Ye1jOE&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FQbQV6Ye1jOE%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/afba1a4a6aa0b45cb414de73a61b29ca/href">https://medium.com/media/afba1a4a6aa0b45cb414de73a61b29ca/href</a></iframe><h3>Expanding Sphero RVR</h3><p>The Sphero RVR is an app-enabled, all-terrain, fully programmable robot. Straight out of the box, the RVR can be driven or programmed, with the Sphero EDU app. And with light sensors, IR sensors, an accelerometer, magnetometer, and a colour sensor, RVR itself is packed with possibility. But this gets even more interesting when you consider that RVR is compatible with Arduino, Raspberry Pi, and micro:bit, as well as Sphero littleBits. This provides two further options: creating an independent “topper” for the RVR platform, or programming the RVR with another micro-controller or micro-computer. And to do this, <a href="https://sdk.sphero.com/">Sphero have released SDK documentation for each board</a>.</p><h3>Programming Sphero RVR with Arduino UNO</h3><p>In this example, I’m using an Arduino UNO compatible board (an Elegoo UNO R3), first to create a “topper” for the RVR, and then secondly to control the actual RVR too. The ultimate aim is to create a robot that navigates small spaces to put out small fires, as part of deeper research into emergency robotics. With the all terrain treads, and the precise control systems that keep RVR in the right direction despite obstacles, the RVR is the perfect choice for this type of project.</p><h3>Step 1: A fire fighting topper</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMDzLI7kIgX4%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMDzLI7kIgX4&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMDzLI7kIgX4%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/fb1b3f5b55bd9f6cd695d2ece2ebcc23/href">https://medium.com/media/fb1b3f5b55bd9f6cd695d2ece2ebcc23/href</a></iframe><p>In the first example, the Sphero RVR and the Arduino components are running separately. This means that the Arduino UNO, flame sensor and fan module are powered by the RVR, and will run independently. The RVR can then be programmed separately or be remotely driven, using the Sphero EDU app.</p><p>To do this, I used a:</p><ul><li>Flame sensor module- the GND pin and VCC pin on the flame sensor connect to the GND and 5V on the breadboard. The A0 pin on the flame sensor connects to the Arduino A0 pin.</li><li>L9110 fan motor module- The GND pin and VCC pin on the flame sensor connect to the GND and 5V on the breadboard, or the GND and 3.3V on the Arduino board. The INA pin connects to Digital Pin 9, and the INB pin to the Digital Pin 8 on the Arduino board.</li><li>breadboard and jumper wires</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*LNozgbybKQcWuzcxcyJ7fA.jpeg" /><figcaption>Setting up the Arduino UNO board and the modules</figcaption></figure><p>As you can see I used a combination of LEGO and tape to hold the sensors in place.</p><p>The flame sensor is programmed to continuously detect for flames, and upon detection, turn on the fan. The fan will remain on while the flame is detected.</p><p>This is the code along with notes:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/920/1*YcSPQrmI4Nl3Cm4tQsqvzw.png" /></figure><h3>Step 2: Programming movement of Sphero RVR with the Arduino UNO</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FJu5JTphF1nM%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DJu5JTphF1nM&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FJu5JTphF1nM%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/7b9a50d9c873e502f4d8f9854b3fa7be/href">https://medium.com/media/7b9a50d9c873e502f4d8f9854b3fa7be/href</a></iframe><p>This was a testing step for me to understand how Sphero RVR’s movement could be programmed with Arduino. Fortunately, the Sphero RVR SDK documentation is incredibly helpful, and I found this process to be easier than expected.</p><p>There are 4 different drive example files for controlling the Sphero RVR with the Arduino UNO, and for this code I used the <a href="https://sdk.sphero.com/docs/how_to/arduino/arduino_how_to/#drive-with-helper">“Drive with Helper” example</a>.</p><p>This program built on the previous version, by simply moving the RVR by 10 degrees to the right, while a flame was not detected, so completing a full circle of the immediate area. While a flame is detected, the movement stops and the fan is switched on to put out the flame, just as before. A more practical way to scan a specific area for flames would be to install a number of flame sensors, to cover each angle at once. But as a testing step, this program was beneficial.</p><p>This is the code:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*s6OC9_lAPd6hevskIFeVTw.png" /></figure><h3>Step 3: Navigation and fire fighting</h3><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FKv8PPMudR5o%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DKv8PPMudR5o&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FKv8PPMudR5o%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="854" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/8e9237d045a9d5200c3f69e17acdff48/href">https://medium.com/media/8e9237d045a9d5200c3f69e17acdff48/href</a></iframe><p>The ultimate aim was to create a robot model that could navigate into small spaces and put out small fires. To do this, I had to add an Arduino module that would help RVR “see” obstacles. An IR obstacle sensor would be effective for this, but I opted for an ultrasonic sensor (or ultrasonic range finder) because I wanted reliable, numerical, information about the distance of obstacles ahead.</p><p>I used a HC-SR04 ultrasonic sensor. The GND and VCC pins on the ultrasonic sensor connect to the GND and 5V on the breadboard. The echo pin connects to Digital Pin 7 and the ping pin connects to Digital Pin 6.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wgFSrA_nX4n-sVEg5kdl0Q.jpeg" /><figcaption>Sphero RVR and Arduino UNO fire fighting robot</figcaption></figure><p>One of the key differences between this program, and Step 2, is that for this version I used the <a href="https://sdk.sphero.com/docs/how_to/arduino/arduino_how_to/#drive-with-heading">“Drive with Heading” example</a>. This has some key differences in the code, including needing to reset the yaw each loop, and not needing to establish drive control in the setup.</p><p>This program detects first for fires, then obstacles, moving forward if neither are detected. On detecting a fire, the fan will be switched on until the fire is put out, and then the rest of the program will resume. If an obstacle is detected within 25cm, RVR will move right (90 degrees). If this way is blocked, RVR will turn to the left (270 degrees), and if this is also blocked, RVR will turn back to face the opposite way (180 degrees).</p><p>I used a function called “searching()” for the code to measure distance, as this is frequently repeated in the same way throughout the program.</p><p>I also used a program called “ fire()” to look for a flame and switch on the fan if necessary, as this is also frequently repeated.</p><p>The code looks like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*tW1tNaWYKMYkfs-WKlVXGQ.png" /></figure><p>It should be noted that this program could be condensed and simplified, if you wanted the robot to simply navigate an open space, with no preference for direction. RVR could simply turn to the right each time an obstacle is detected. This would not necessarily be effective for small spaces, as when faced with an obstacle, turning right and then faced with another obstacle, RVR would turn right again, leaving the space without fully exploring to the left.</p><p>The code for the simplified version would look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/848/1*naOwKrM1TayP9kUlxlrCHw.png" /></figure><h3>Overall…</h3><p>…the Sphero RVR is a versatile and flexible robotic platform that could be used for a wide range of purposes in STEM education, and for hobbyists. It’s easy to get going with the SDK documentation and the programming possibilities are endless. While I don’t think that Sphero RVR will have a future career in the fire service, the fire fighting robot program was successful on the whole.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=1a4220ec581f" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Show images and play sound with Vector Robot]]></title>
            <link>https://medium.com/@ashbn/show-images-and-play-sound-with-vector-robot-39f46b8a0b09?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/39f46b8a0b09</guid>
            <category><![CDATA[vector-robot]]></category>
            <category><![CDATA[programming]]></category>
            <category><![CDATA[star-wars]]></category>
            <category><![CDATA[python]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Sun, 28 Jun 2020 16:26:48 GMT</pubDate>
            <atom:updated>2020-06-28T16:50:49.606Z</atom:updated>
            <content:encoded><![CDATA[<iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2Fr3cHA_0rCSo%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3Dr3cHA_0rCSo&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2Fr3cHA_0rCSo%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/2c899d03527b125e659e270d57be3e02/href">https://medium.com/media/2c899d03527b125e659e270d57be3e02/href</a></iframe><p>Program Vector robot to show images and play sound files with the SDK.</p><p>Using Vector’s screen to show images, and the sound system to play sound files and music, are just two ways to program Vector to deliver a custom message. For any occasion, like Star Wars Day, this can be a fun way to celebrate. But how can you use the SDK for this?</p><h4>The current status of Vector Robot</h4><p>Vector robot, the home assistant robot with personality, was developed by Anki and released in October 2018. The future of this little robot was uncertain when Anki collapsed in April 2019. Since then Digital Dream Labs (DDL) have taken the helm, and after an incredible Kickstarter campaign, they have already began to add useful features and updates to this personable robot companion. With a new SDK system promised for development, it’s important to note that this guide uses the old Anki Vector SDK (version 6.0) which you can download <a href="https://developer.anki.com/vector/docs">here</a>.</p><h4>SDK “tutorials” folder</h4><p>Once you have the SDK downloaded and set up with your Vector, the example files are a great place to start. First, navigate to the “tutorials” folder. Here you can find example codes for how to program Vector to speak, move, play animations, interact with the environment, show images and play sounds. If you are new to python and programming, you might want to start at the top and work through some of the easier examples. <a href="https://www.hexnub.com/programming-anki-vector-using-the-sdk/">This is a great guide to help you get started.</a></p><h4>Using the SDK “face_image” example</h4><p>Example 06 is the “face_image” example code. In the tutorials folder, right click on this file and select “Edit with IDLE” to open the program in the python programming environment. This is an example program that will display an image of Cozmo (Anki’s other robot) on Vector’s face. It should look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/981/1*UEnRh1OwaPCbMQUPVwn0HQ.png" /><figcaption>06_face_image example- Vector SDK</figcaption></figure><p>The red notes at the top are not relevant to the code. But the other red notes provide important pointers, indicating the purpose of some of the lines of code. Essentially this program will:</p><ol><li>Lift Vector’s head to show the screen more effectively</li><li>Find the image file by following the specified file path</li><li>Load the image</li><li>Display the image for a specific duration</li></ol><h4>Displaying a different image on Vector’s face</h4><p>To change which image is displayed on Vector’s face, you need to follow these steps:</p><ol><li>Find the image you want</li><li>Resize this image to 184 x 96 pixels. Vector will not load any images that do not fit this size criteria and your code will return an error.</li><li>Save the resized image in the “face_images” folder (above the “tutorials folder in the SDK example files)</li><li>Change “cozmo_image.jpg” within the example code to the file name of your new chosen image. Make sure you use the perfect spelling.</li><li>Save your modified program in the same “tutorials” folder with a new descriptive name.</li><li>Select “Run” and the “Run Module” to see your program in action on Vector.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/878/1*d31tSPme_XN4jmIma-4uXw.png" /><figcaption>This will display the image of R2D2 instead of Cozmo. I also removed the red notes for tidier code, and removed the print instruction as it was not needed.</figcaption></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*wt-MdHAIq0mKuMR_gRLbXQ.jpeg" /><figcaption>Vector displaying R2D2 image</figcaption></figure><h4>Using the SDK “play_audio” example</h4><p>The SDK “play_audio” can be difficult to get started with. When you open the example code, it should look like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*OXU4K786GBUIOkBVCvoElQ.png" /><figcaption>10_play_audio example- Vector SDK</figcaption></figure><p>As you can see, you will need to navigate to the github web address to download the sound files. These should be saved into the same “tutorials” folder as the code itself. With those downloaded, you should be able to get Vector to play the sound file by hitting “Run” and then “Run module”.</p><h4>Playing a different sound file through Vector</h4><p>To change the sound file that Vector plays you will need to:</p><ol><li>Find the sound file that you want</li><li>Convert this to the right format. This needs to be a .wav file, that is 16bits, single channel, with a sample rate of 8000–16025 samples/sec. I find <a href="https://twistedwave.com/online">TwistedWave Online</a> is a helpful free tool for resampling any music files.</li><li>Save the converted sound file to the “tutorials” folder. It needs to be the same folder as the code file itself.</li><li>Change the name of the wav file in the code, to the name of your new music or sound file.</li><li>Change the volume if you want to, by changing the value after the name of the sound file. It is set to 75 but it will accept any value between 0 and 99.</li><li>Save the program with a different and descriptive name in the “tutorials” folder.</li><li>Hit “Run” and then “Run Module” to test your program.</li></ol><p>Optional: Now you know what each part is for, you can also remove the red lines of notes for a tidier and cleaner program.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/555/1*N_X7qHFmksPCInzJm2goUA.png" /><figcaption>My program will now play R2D2 sound effects.</figcaption></figure><p>You can also play music, not just sound effects. Like this video for a Valentines Day message:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FU9rHD2rWlcA%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DU9rHD2rWlcA&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FU9rHD2rWlcA%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/ee3940d81ded2b724d2dd8a125c684a1/href">https://medium.com/media/ee3940d81ded2b724d2dd8a125c684a1/href</a></iframe><h4>Program Vector robot to display images and play sound at the same time</h4><p>To combine your new sound and image files together, and display an image while playing a sound file, you just need to copy the play audio line of code from your new sound file, and paste this into the image code. Like this:</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/887/1*jRBdBcIC3ZXApi7uExTpfg.png" /><figcaption>Now Vector will raise his head, display an image of R2D2 for 20 seconds, while playing R2D2 sound effects.</figcaption></figure><p>One thing to consider is the length of time that you want the image to be displayed, or the duration. If you want the image to show the entire time that your sound or music file is playing, you will want to set the duration to the same length.</p><p>You will also need to remove the delay, as you don’t want the program to pause, you want it to continue so that the sound file will play while displaying the image.</p><p>This is the code in action:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMdDK_AXoiPQ%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMdDK_AXoiPQ&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMdDK_AXoiPQ%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/a183083da58f6da8efa17bf5acdb83a4/href">https://medium.com/media/a183083da58f6da8efa17bf5acdb83a4/href</a></iframe><h4>Programming a custom response to a verbal command</h4><p>To take this one step further, you can make sure that your Vector acts as normal, until a specific user command is issued, like “Hey Vector…good job” or “Hey Vector…I love you”. At this point, Vector can then deliver his custom message, complete with music and screen images. For example, this star wars program is triggered when the user says “The force be with you”, which registers as a user intent “imperative praise”. To find out more about user intents, check out this <a href="https://medium.com/@ashbn/anki-vector-sdk-a-robot-with-attitude-618fbb081802">guide</a>.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*Me3NWuTzhtnLPHLm-MpZug.png" /><figcaption>This is the code for the video shown at the top of this page.</figcaption></figure><p>With these examples there is no limit to the messages your Vector can deliver. Like this <a href="https://www.prnewswire.com/news-releases/simon-philips-former-president-of-marvel-entertainment-international-joins-robotics-maker-digital-dream-labs-board-of-advisors-301035918.html">hand washing video</a> for example, or this Christmas song jukebox:</p><iframe src="https://cdn.embedly.com/widgets/media.html?src=https%3A%2F%2Fwww.youtube.com%2Fembed%2FMwJWiJIk54o%3Ffeature%3Doembed&amp;display_name=YouTube&amp;url=https%3A%2F%2Fwww.youtube.com%2Fwatch%3Fv%3DMwJWiJIk54o&amp;image=https%3A%2F%2Fi.ytimg.com%2Fvi%2FMwJWiJIk54o%2Fhqdefault.jpg&amp;key=a19fcc184b9711e1b4764040d3dc5c07&amp;type=text%2Fhtml&amp;schema=youtube" width="640" height="480" frameborder="0" scrolling="no"><a href="https://medium.com/media/2c9043a011e737543582105d7c7f9266/href">https://medium.com/media/2c9043a011e737543582105d7c7f9266/href</a></iframe><p>Check it out and see what you can create. Let me know what you think in the comments!</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=39f46b8a0b09" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Programming a turn based role playing game with Sphero BOLT]]></title>
            <link>https://medium.com/@ashbn/programming-a-turn-based-role-playing-game-with-sphero-bolt-bf2d4dc878ba?source=rss-498ada985ef0------2</link>
            <guid isPermaLink="false">https://medium.com/p/bf2d4dc878ba</guid>
            <category><![CDATA[sphero-bolt]]></category>
            <category><![CDATA[javascript]]></category>
            <category><![CDATA[sphero]]></category>
            <category><![CDATA[dungeons-and-dragons]]></category>
            <category><![CDATA[role-playing-games]]></category>
            <dc:creator><![CDATA[AshbotandSparki]]></dc:creator>
            <pubDate>Wed, 22 Jan 2020 13:37:56 GMT</pubDate>
            <atom:updated>2020-01-22T13:37:56.435Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*cUl5rMOFJk9Dmcw3vCrtDw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zulnQXphTpt_GzwslG0GBA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7gH2EHxfNiiLGn-4GXUp8Q.jpeg" /><figcaption>Create your own D&amp;D game night with Sphero BOLT</figcaption></figure><p>One of the coolest aspects of the Sphero BOLT is the complete programming freedom. From maze solving to quizzes, you can create endless codes for this app enabled robot ball. And the programmable LED matrix is also a nice touch, adding to the level of creativity on offer. So I decided to push this to the limit, aiming to create an immersive, turn based RPG.</p><h3>Sphero BOLT turn based RPG mechanics</h3><p>To create this turn based role playing game, there were a number of game mechanics that had to be established. These include:</p><ul><li>Speech- To deliver the story-line and quests, and indicate points where player interaction is necessary, the “await speak” command had to be used. As a result, all of the story-line is delivered through speech, in a similar way to table top RPG’s.</li><li>Shaking BOLT- Velocity is the easiest of BOLT’s many sensors to measure, for all players, without needing to calibrate the robot first. This means that shaking Sphero BOLT is a simple way for users to make a choice and play the game.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*HT3drONhgzprB3snsH5T0A.png" /><figcaption>By measuring the combined velocity of the Sphero BOLT, it is easy to establish when the robot has been shaken.</figcaption></figure><h3>Getting started</h3><p>To get started with this project, there were a number of things to decide on, including:</p><ul><li>The type of setting- I opted for the typical medieval castle/dungeon as the setting for the RPG.</li><li>The class types- In this game you can choose between two classes, a fighter and a mage.</li><li>The quest line- I wanted to keep this RPG fairly simple, due to the limitations of the device. So I opted for a simple dungeon style quest line where the hero (you) has to find a lost item by moving through the dungeon, slaying monsters.</li></ul><p>That said, the first thing the user is asked to do as part of this program is to select a class.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*7gH2EHxfNiiLGn-4GXUp8Q.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-8i5AyvgLeD7cZZZ6VZY0Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/461/1*0cy19rE4BDCzSrTST0uqeg.png" /><figcaption>(Finished ice staff animation for the class type) JavaScript code for choosing a class. (Ice staff in app design)</figcaption></figure><p>From there, the player is given some information regarding their starting weapon, and the quest. This will depend on the chosen class, as each has a slightly different mission, and will face different enemies. This is all established in the many variables within each class function.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*DBb3DVEohel1uUd07iRcGw.png" /><figcaption>The class functions allow a number of variables to be assigned different values, which will be used to determine the course of the game.</figcaption></figure><h3>Setting different functions</h3><p>For a game of this size to work, different functions were necessary. As well as functions for defining each character class, as shown above, other functions needed include:</p><ul><li>Each quest line- when adding in a checkpoint system the quest lines had to be divided into different functions for each part of the quest line. This includes the final part of the quest, where the quest item is retrieved.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*B3dgsboWOVUK38ghab77vw.png" /></figure><ul><li>Opponent choice- a function to establish which opponent you are fighting each time, is important so that the player fights a range of opponents. There are 8 different opponents altogether.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*G_yr-xMP5N8xqMfCwe2gUg.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/672/1*HykJymouYp2UUYVjt9QJ9Q.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/461/1*NzolfSs5-9mK8UvbFeYhBg.png" /><figcaption>(Finished Skeleton animation) JavaScript code for determining which opponent you will battle. (In app Skeleton animation design)</figcaption></figure><ul><li>Health points- storing the health points in a function helps to keep track of them during each battle. It was also necessary to create separate function for the health points of the opponent.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qGtUs6dpP2GLhlVRnDsKPQ.png" /></figure><ul><li>Attack damage- a function to establish the damage caused by each attack is needed, as this will be called multiple times throughout the game. A separate function is also important to establish the damage caused by the opponents. These functions link closely with the health point functions too.</li><li>Checkpoint- added in later on in the development stages, the checkpoint function allows the player to restart at an earlier point, after losing all of the health points.</li></ul><figure><img alt="" src="https://cdn-images-1.medium.com/max/445/1*87PORVtG35ILcNxFeYbtaA.png" /></figure><h3>Developing a combat system for a turn based RPG</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*zulnQXphTpt_GzwslG0GBA.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*YVuNy9oDzNYYipyD_HZQYA.jpeg" /><figcaption>Fighting a fire troll.</figcaption></figure><p>Creating a combat system was complicated and there were several different iterations of the combat system before the final version was selected. These were the stages of development:</p><ol><li>A combat system where the user would shake BOLT to “roll the dice”, shown on the LED matrix, which would then determine the attack damage. This system made an effective link to the D&amp;D style RPG games, but it did involve a lot of shaking.</li><li>A combat system where the attack damage is based on a random integer, decided by the program, rather than the “roll the dice” animation. This was more straightforward for the user, but it also felt a little less immersive.</li><li>A combat system where the player can choose between 2 different attack options, by shaking the BOLT. Different attacks result in different amounts of random damage. The first attack option will deal between 20 and 30 points of damage. The second attack option has a 20% chance of failing completely, a 60% chance of causing 30 to 40 points of damage, and a 20% chance of activating a power attack, causing 40 to 50 points of damage.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VrquucwD3-2aEFlUS2WXuA.png" /><figcaption>Being able to choose an attack is an important part of any turn based RPG.</figcaption></figure><p>In the end I opted for choice number 3, but I also added a “block” option for attacks too, which allows you to block the next, incoming attack. However, this will not deal any damage to your opponent.</p><p>Enemy attacks are set to a random integer, with boss enemies dealing larger amounts of damage than common dungeon enemies.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*uA5TDfFIyf2zpynzapInew.png" /></figure><h3>Visual feedback for damage caused</h3><p>Providing feedback for the amount of damage caused by the players attack, and by the enemy attack, was also a necessary consideration. Initially I set this up as a spoken review, after each attack, but between choosing the attack, and announcing the enemy attack, there seemed to be too much speech. Instead a visual feedback system is used. This shows the enemy health as measured in a green heart animation, and the players health as measured in a red heart. This provides instant visual recognition, so you can plan and strategise your next attack accordingly.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*MiuPIFbpkMuVMpwvrHE-EQ.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/445/1*Y5iMlcb-xRUedDOqJcXIGg.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qGtUs6dpP2GLhlVRnDsKPQ.png" /><figcaption>(Finished animation for player health) In app player health design. (JavaScript code for determining the player health during battle)</figcaption></figure><h3>Checkpoints</h3><p>Another aspect necessary to include for any RPG are checkpoints. One of the limitations in creating a turn based RPG for the Sphero BOLT is that progress cannot be saved between games. So once the program ends, any data is lost. However, while the program is still running, it is possible to go back to an earlier point. Checkpoints are used to save progress at various points, so that if the player is unlucky in battle, they don’t have to replay too many enemies again, to return to where they were before losing health points.</p><p>To do this, the quest stages have been broken up into different functions. This means that the player can restart at the beginning of the latest dungeon chamber, without needing to replay the earlier part of the game.</p><h3>Healing</h3><p>After each dungeon chamber the user can heal with a healing potion. This is not optional, as without healing you will be sure to lose health in the following chamber. Each healing point is the same for each class and quest line and works as a repeated function for a more efficient program. This restores the hp variable to 100 points.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/455/1*lNS3vNuWChSljTocqNOAcw.png" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*jj3LizoI_oMU9pE4Sf5KVw.jpeg" /></figure><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*M4L70WFN5Km3p85Uhyk1gg.png" /><figcaption>(Healing potion design in app) Finished Healing potion animation. (JavaScript code for healing points)</figcaption></figure><h3>Finishing touches</h3><p>Animation and sound effects were added once all the mechanics were in place (it was actually at this point that the thief quest line had to be removed, in order to run the program on Sphero Edu). The animations displayed on the LED matrix provide visual interest in line with the story which is told through speech. The sound effects are used for dramatic effect, and to create a more immersive player experience.</p><h3>The finished project</h3><p>The finished project is a turn-based adventure quest that involves shaking Sphero to make choices and take action. The player will need to listen to the emerging story line, as they delve into the dungeon and battle monsters in the dark.</p><p>There are an unlimited number of lives, but each time you lose health points you will need to replay the game from the last checkpoint. It can take between 4 and 10 minutes to complete one play through of the game, realistically. This means you can play through easily, without Sphero BOLT running out of charge. And if you choose a different class, you can then embark on a different quest the second time around.</p><p><a href="https://edu.sphero.com/remixes/6596077">This is a link to the Sphero EDU program</a>.</p><h4>If you want access to the direct JavaScript code just let me know :)</h4><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=bf2d4dc878ba" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>