<?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 SW on Medium]]></title>
        <description><![CDATA[Stories by SW on Medium]]></description>
        <link>https://medium.com/@s.wakayama?source=rss-364ed87db857------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*mk2vfAiw78wKzdRC</url>
            <title>Stories by SW on Medium</title>
            <link>https://medium.com/@s.wakayama?source=rss-364ed87db857------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Mon, 16 Nov 2026 15:37:09 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@s.wakayama/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[Strategic Information Gathering: Powering Autonomous Drones with Monte Carlo Tree Search]]></title>
            <link>https://medium.com/helios-techblog/strategic-information-gathering-powering-autonomous-drones-with-monte-carlo-tree-search-78c51072099a?source=rss-364ed87db857------2</link>
            <guid isPermaLink="false">https://medium.com/p/78c51072099a</guid>
            <dc:creator><![CDATA[SW]]></dc:creator>
            <pubDate>Sun, 06 Jul 2025 07:24:48 GMT</pubDate>
            <atom:updated>2025-07-28T07:40:50.526Z</atom:updated>
            <content:encoded><![CDATA[<p>Satellite remote sensing is a powerful tool for obtaining a bird’s-eye view of vast areas. Publicly available data such as that from <a href="https://landsat.gsfc.nasa.gov/satellites/landsat-8/">Landsat</a> and <a href="https://www.esa.int/Applications/Observing_the_Earth/Copernicus/Sentinel-2">Sentinel-2</a> allows us to capture large-scale changes in vegetation and water quality, etc. However, the spatial resolution of these platforms is limited to around 10–30 meters. Consequently, even if signs of an anomaly are detected in an observed area, this resolution is insufficient to clarify the details.</p><blockquote>For instance, while water discoloration in coastal areas can be detected from satellite images, it is difficult to discern whether it signifies the degradation of seagrass beds — which would have a serious impact on the surrounding ecosystem — or is merely a temporary phenomenon. Furthermore, details such as the accumulation of microplastics on shorelines or fine cracks in infrastructure remain indiscernible.</blockquote><p>Centimeter-grade, high-resolution commercial satellites do exist, but using them to cover extensive areas is not cost-effective. Moreover, the inherent weakness of optical sensors — being obstructed by clouds — persists. So, the use of drones equipped with high-resolution cameras is anticipated as a means to resolve this satellite-indicated “uncertainty” and enhance situational awareness. However, drones have limited flight times. While manual operation by a specialized pilot is an option, the associated mental burden and personnel costs are not insignificant.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*XqkV2bttXpEdKdWqES9mQw.png" /><figcaption>Leveraging heterogeneous data from diverse sources is key to enhancing situational awareness.</figcaption></figure><p>Therefore, this post introduces <a href="https://arxiv.org/abs/2103.04931"><em>Monte Carlo Tree Search (MCTS)</em></a> as one powerful method for enabling a drone to efficiently and autonomously gather information in an unknown environment.</p><h3>Overview of Monte Carlo Tree Search</h3><p>Monte Carlo Tree Search is an algorithm famously adopted by “<a href="https://www.nature.com/articles/nature16961">AlphaGo</a>” — renowned for defeating top professional players in the game of Go — and has garnered significant attention in the fields of game AI and, more recently, robotics. This algorithm is known to be powerful for complex decision-making problems where the goal is to determine the best option from a long-term perspective among numerous choices.</p><blockquote>Such decision-making problems can often be modeled with a mathematical framework known as the <a href="https://www.jstor.org/stable/24900506">Markov Decision Process (MDP)</a>, where selecting an “action” in a given “state” leads to a probabilistic transition to a subsequent “state” and the receipt of a “reward.” For such problems, MCTS is an exceptional approach that, instead of simulating a brute-force search of all conceivable long-term action plans, efficiently identifies the most promising options.</blockquote><p>Similar to how a brilliant Go player mentally simulates countless move sequences to find the optimal move, MCTS utilizes a tree-like data structure called a “search tree” to organize its thought process, ultimately determining the optimal choice by executing four key steps (1. Selection, 2. Expansion, 3. Rollout/Simulation, 4. Backpropagation) hundreds or thousands of times at high speed within limited computational resources.</p><p>In the context of the information-gathering task to reduce environmental uncertainty, these steps can be described as follows:</p><ol><li><strong>Selection:</strong> First, an action is selected from the search tree built upon previous simulation results. At this stage, a balance is carefully struck between actions already known to have high expected information gain and unexplored actions with untapped potential.</li><li><strong>Expansion:</strong> Next, unexplored action options (e.g. the drone’s next potential waypoints) are added to the search tree beyond the selected action.</li><li><strong>Rollout/Simulation:</strong> Then, from the newly added action, the drone is virtually flown to the end of a predefined planning horizon. The total amount of information gain (i.e. reward) that can be ultimately obtained in this virtual mission is then evaluated.</li><li><strong>Backpropagation:</strong> Finally, the result (total information gain) from the simulation is fed back to the traversed branches of the tree. In other words, the value estimate is updated to reflect that selecting a particular action can be expected to yield a certain level of outcome.</li></ol><figure><img alt="" src="https://cdn-images-1.medium.com/max/860/1*SgIRVcwHMnwihQdRDEEgTw.png" /><figcaption>From <a href="https://www.google.com/imgres?q=monte%20carlo%20tree%20search&amp;imgurl=https%3A%2F%2Fupload.wikimedia.org%2Fwikipedia%2Fcommons%2Fc%2Fcb%2FMCTS-diagram.svg&amp;imgrefurl=https%3A%2F%2Fcommons.wikimedia.org%2Fwiki%2FFile%3AMCTS-diagram.svg&amp;docid=ePjmDomAMDQmQM&amp;tbnid=gAMQQWo2HCnnXM&amp;vet=12ahUKEwirs6G4w5GOAxVvj68BHUPTO5AQM3oECGoQAA..i&amp;w=918&amp;h=261&amp;hcb=2&amp;ved=2ahUKEwirs6G4w5GOAxVvj68BHUPTO5AQM3oECGoQAA">Wikimedia Commons</a></figcaption></figure><p>As this cycle is repeated, the search tree intelligently grows toward statistically promising sequences of actions. Then, when the search resources are exhausted, the action deemed most promising at that moment is selected and executed as the drone’s next real-world move.</p><h3>Drone Behavior with MCTS</h3><p>Let’s look at an example of how a simulated drone using this algorithm behaves in a coastal information-gathering task. In this scenario, environmental uncertainty is represented by a heatmap overlaid on a satellite image, where its degree is shown by a color gradient ranging from pink (low) to white (high).</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*Im1RuJAvNQLPuM1CWG5k_g.gif" /></figure><p>In the GIF above, you can see that even without any direct flight path commands, the drone acts as if it has its own volition, autonomously generating a path that actively explores areas of high uncertainty (the white areas in the figure). This is because MCTS is “planning ahead” to find a path that can resolve more uncertainty over the long term. Even a move that seems like a detour at first glance is a strategic decision to maximize information gain within the planning horizon.</p><p>In the field of remote sensing data analysis, research and development often focus on how to derive insights from passively acquired data. At <a href="https://www.helios-rs.com/?utm_source=Helios-Medium&amp;utm_medium=Medium&amp;utm_id=250728-01">Helios</a>, in addition to this, we also<em> engage in data acquisition itself </em>— determining how to collect data that is truly useful to users, as demonstrated in the example above.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=78c51072099a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/helios-techblog/strategic-information-gathering-powering-autonomous-drones-with-monte-carlo-tree-search-78c51072099a">Strategic Information Gathering: Powering Autonomous Drones with Monte Carlo Tree Search</a> was originally published in <a href="https://medium.com/helios-techblog">Helios-techblog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Unpacking the Four Layers of Autonomous Mobile Robot Systems]]></title>
            <link>https://medium.com/helios-techblog/unpacking-the-four-layers-of-autonomous-mobile-robot-systems-eeadc7bb326f?source=rss-364ed87db857------2</link>
            <guid isPermaLink="false">https://medium.com/p/eeadc7bb326f</guid>
            <category><![CDATA[autonomous-system]]></category>
            <category><![CDATA[robotics]]></category>
            <dc:creator><![CDATA[SW]]></dc:creator>
            <pubDate>Thu, 19 Jun 2025 07:54:28 GMT</pubDate>
            <atom:updated>2025-07-28T07:40:07.678Z</atom:updated>
            <content:encoded><![CDATA[<p>Autonomous mobile robots are becoming indispensable for our lives because they can perform tedious, repetitive, and hazardous tasks on our behalf. These include crucial operations such as conducting large-scale environmental monitoring, exploring unknown territories like outer space and deep caves, and assessing a situation in the immediate aftermath of a disaster. But how do these machines translate a complex mission into autonomous actions? Typically, this is achieved by processing information sequentially from top to bottom through a four-layer architecture as illustrated below. In this post, we will dive into each of these four essential layers to provide you a clearer picture of how this remarkable technology works.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/403/1*AnKL6Cyg_RDhzJozU9IPng.png" /><figcaption>Four layers of autonomous mobile robot systems</figcaption></figure><h3>1. Decision-Making Layer</h3><p>At the very top of the hierarchy sits the Decision-Making layer. This layer acts as the “brain” of autonomous mobile robots. Its primary purpose is to make high-level strategic decisions to accomplish a given mission. For instance, in a bridge inspection mission, this layer determines a long-term, global strategy, such as which area of the bridge to survey next. The goals can range from “map this entire area efficiently” to “find and monitor all intruders in this facility” or “inspect all turbines in this wind farm.”</p><p>When a mission involves information gathering, the problem is often known as Informative Path Planning (IPP). The objective of this problem is to plan a path that maximizes the amount of information the robot can gather, all while operating under the constraints of its limited resources, such as battery life or time. To solve this, several strategies can be employed.</p><p>One classic and intuitive strategy is the Next-Best-View (NBV) approach. As the name suggests, it works as follows:</p><ol><li>A robot generates a set of potential next viewpoints it could move to</li><li>It then calculates the expected amount of new information it would gain from each candidate viewpoint</li><li>The viewpoint offering the most information is selected as the “next-best-view”, and the robot moves there</li><li>This process is repeated until the mission is complete. A notable implementation is detailed in the paper, <a href="https://ieeexplore.ieee.org/document/7487281"><em>Receding Horizon “Next-Best-View” Planner for 3D Exploration</em></a> (ICRA 2016).</li></ol><p>For situations that require handling uncertainty more rigorously and making non-myopic decisions, an IPP problem can be formulated as a Partially Observable Markov Decision Process (POMDP). This mathematical framework allows a robot to plan actions that are optimal in the long run. However, solving a full-scale POMDP is often analytically intractable for large state and observation spaces. Therefore, online sampling-based algorithms such as <a href="https://papers.nips.cc/paper_files/paper/2010/hash/edfbe1afcf9246bb0d40eb4d8027d90f-Abstract.html">POMCP</a>, <a href="https://arxiv.org/abs/1709.06196">POMCPOW</a>, <a href="https://arxiv.org/abs/1609.03250">DESPOT</a> are used to find approximate solutions.</p><blockquote><strong>Practical Considerations: </strong>Nevertheless, their computational demands can make them impractical for deployment on real-world robots. Hence in many scenarios, simpler and more cost-effective strategies like a pre-defined Lawnmower Pattern or the NBV approach are sufficient. However, for special cases, such as responding to a disaster where no reliable prior map exists, the robustness of a POMDP-based planner could prove highly effective.</blockquote><p>To achieve other types of goals beyond information gathering, this layer might employ different frameworks like <a href="https://robohub.org/introduction-to-behavior-trees/">Behavior Trees (BT)</a> or <a href="https://robohub.org/task-planning-in-robotics/">Planning Domain Definition Language (PDDL)</a>, but we will not go into the details of those in this post.</p><h3>2. Path Planning Layer</h3><p>Once the Decision-Making layer has set a high-level goal (e.g. “go to location B”), the Path Planning layer is responsible for determining a collision-free, geometric route from the robot’s current position (Point A) to that destination (Point B). This layer takes into account a map of the environment, which includes static obstacles like buildings and walls, as well as potentially dynamic obstacles like other robots or people.</p><p>The key output of this layer is a geometric path — represented as a set of points or a curve — that contains no timing information. Its main purpose is to provide a safe route.</p><p>A variety of well-known algorithms are used to accomplish this. Famous examples include <a href="https://en.wikipedia.org/wiki/A*_search_algorithm">A*</a>, which is excellent for finding the shortest path in a grid-based map, as well as sampling-based methods like <a href="https://en.wikipedia.org/wiki/Rapidly_exploring_random_tree">RRT (Rapidly-exploring Random Trees)</a> and its variants, and <a href="https://en.wikipedia.org/wiki/Probabilistic_roadmap">PRM (Probabilistic Roadmap)</a>, which are highly effective in complex, high-dimensional spaces.</p><p>The path generated here is safe, but it is not yet ready for motors to execute directly, as it does not the robot’s dynamic constraints like velocity or acceleration. That task is left to the next layer.</p><h3>3. Trajectory Planning Layer</h3><p>The path from the previous layer is essentially a series of connected points. The Trajectory Planning layer takes this path and refines it into a smooth, time-parameterized trajectory that the robot can physically follow. This means the output is not just <em>where</em> the robot should go, but <em>when</em> it should be at each point along the path.</p><p>This layer is crucial because it accounts for the robot’s physical limitations, such as maximum velocity, acceleration, and jerk. By optimizing the trajectory, we ensure the robot’s movement is not only smooth and efficient but also within the bounds of what its hardware can handle. This prevents jerky movements, reduces energy consumption, and minimizes wear and tear on the robot’s mechanical parts.</p><h3>4. Control Layer</h3><p>At the lowest level, we have the Control layer. This layer is the final link between the digital plan and the physical world. It takes the optimized trajectory from the layer above and translates it into low-level commands for the robot’s actuators — the motors that drive the wheels or propellers.</p><p>The controller’s job is to ensure the robot follows the desired trajectory as closely as possible in real-time. It uses feedback from sensors (like GPS, IMU, and encoders) to continuously measure the robot’s actual state (position, velocity, etc.) and compares it to the desired state from the trajectory. If there’s a discrepancy, the controller calculates and sends corrective commands to the motors. PID controllers are a common and effective choice for this task.</p><h3>Conclusion and Our Focus at Helios</h3><p>In summary, the autonomy of a mobile robot is achieved through a stack of four interconnected systems: the high-level <strong>Decision-Making layer</strong> for strategic goals, the <strong>Path Planning layer</strong> for finding a safe route, the <strong>Trajectory Planning layer</strong> for generating a smooth and feasible trajectory, and the <strong>Control layer</strong> for executing the motion.</p><p>At <a href="https://www.helios-rs.com/?utm_source=Helios-Medium&amp;utm_medium=Medium&amp;utm_id=250728-01">Helios</a>, we apply advanced <strong>Decision-Making </strong>to create large-scale, high-resolution maps cost-effectively. We solve the core challenge of where to collect high-resolution data by formulating it as an IPP problem. Our service enables automated systems to efficiently explore and gather the most valuable data possible, moving beyond pre-set plans.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=eeadc7bb326f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/helios-techblog/unpacking-the-four-layers-of-autonomous-mobile-robot-systems-eeadc7bb326f">Unpacking the Four Layers of Autonomous Mobile Robot Systems</a> was originally published in <a href="https://medium.com/helios-techblog">Helios-techblog</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>