<?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 Aranzapachecosuarez on Medium]]></title>
        <description><![CDATA[Stories by Aranzapachecosuarez on Medium]]></description>
        <link>https://medium.com/@aranzapachecosuarez?source=rss-8eea93e01a18------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*Z4fCMlwYjxNxpvxH</url>
            <title>Stories by Aranzapachecosuarez on Medium</title>
            <link>https://medium.com/@aranzapachecosuarez?source=rss-8eea93e01a18------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Thu, 19 Nov 2026 05:37:32 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@aranzapachecosuarez/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[The Next Move: How Perseverance Chooses Its Path on Mars]]></title>
            <link>https://medium.com/@aranzapachecosuarez/the-next-move-how-perseverance-chooses-its-path-on-mars-9cac13451ecd?source=rss-8eea93e01a18------2</link>
            <guid isPermaLink="false">https://medium.com/p/9cac13451ecd</guid>
            <category><![CDATA[robots-think]]></category>
            <category><![CDATA[robotics-technology]]></category>
            <category><![CDATA[automation]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[robotics-automation]]></category>
            <dc:creator><![CDATA[Aranzapachecosuarez]]></dc:creator>
            <pubDate>Wed, 28 Oct 2026 01:26:36 GMT</pubDate>
            <atom:updated>2026-10-28T01:27:57.432Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*hcBeBXXkTsATECA_mXIAsA.png" /><figcaption>AI-generated illustration of rover navigation on Mars. The glowing paths are conceptual, not actual AutoNav data.</figcaption></figure><h4>Inside the maps, competing routes, and safety checks behind a rover’s autonomous decisions.</h4><p>Imagine a robot entering a damaged building. A corridor ahead is blocked. One way around it is steep; another takes longer and uses more battery. The connection to its operator is weakening.</p><p>What should it do next?</p><p>“Use AI” sounds like an answer until you ask what the robot actually calculates. Does it choose the shortest route? The safest? What happens when those are different paths?</p><p>That is the question that interests me about autonomous robots: how a goal becomes a physical action when getting it wrong has consequences. To explore it, I’m going to follow one real system: Perseverance’s autonomous navigation on Mars.</p><p>Its decisions become easier to understand when we separate three things: <strong>what looks promising, what passes the safety checks, and what the robot has enough information and time to attempt.</strong></p><h3>Why the next move cannot wait for Earth</h3><p>NASA explains that commands take approximately <a href="https://science.nasa.gov/blog/speaking-percys-language-how-to-talk-so-perseverance-will-listen/">3–22 minutes to reach Perseverance</a>, depending on the planets’ positions [1]. A response takes another journey. That rules out the kind of instant correction we expect when controlling a remote toy.</p><p>Humans still set the broader route and activities. <a href="https://www.nasa.gov/solar-system/nasas-self-driving-perseverance-mars-rover-takes-the-wheel/">AutoNav handles local navigation</a>, building terrain maps and finding routes around hazards [2]. The rover therefore has freedom inside a mission defined by people.</p><p>My starting framework is <strong>Perceive → Reason → Act → Feedback</strong>. But naming a loop doesn’t explain a decision. To do that, we need to look inside the “reason” step.</p><p>Figure 1. Navigation-camera mosaic of Perseverance and the terrain at Snowdrift Peak. The rocks make the challenge of local route planning visible. Image: NASA/JPL-Caltech. <a href="https://www.jpl.nasa.gov/news/autonomous-systems-help-nasas-perseverance-do-more-science-on-mars/">Original image and explanation</a> [3].</p><h3>Perceive: turn pictures into terrain</h3><p>Perseverance’s navigation architecture uses stereo images and visual odometry: estimating motion from images helps combine observations into an evolving elevation map. Its Vision Compute Element accelerates image processing, and “thinking while driving” lets navigation computations overlap with motion. These are documented design features in <a href="https://www-robotics.jpl.nasa.gov/media/documents/AAS_2020_mobility_mmc_v10.pdf">Michael McHenry and colleagues’ engineering paper</a> [4].</p><p>Think about the difference between a photograph and a surface you could stand on. A photograph shows appearance; a terrain model lets a planner reason about relative heights. A route that looks open from one viewpoint might still put a wheel into an awkward position.</p><p>The map is therefore an input to a decision, rather than a decision by itself. “I can see a gap” and “my body can safely fit through that gap” are different statements.</p><h3>Reason: compare routes before spending expensive computation</h3><p>The <a href="https://arxiv.org/pdf/2011.06022">ENav baseline described by Neil Abcouwer and colleagues</a> ranks paths using maneuver time, sampled terrain cost, and estimated remaining cost to the goal, calculated using Dijkstra’s algorithm. Terrain costs can combine tilt, roughness, and traversal time; prohibited regions receive infinite cost. Their experimental tree contains 14 initial turn options followed by two sets of 11 three-meter arcs: <strong>14 × 11 × 11 = 1,694 candidate paths</strong>, looking about six meters ahead. These are reported experimental settings, rather than universal flight settings [5].</p><p>Here is my shorthand for that initial ranking, not a quoted flight-software equation:</p><p>Ranking cost ≈ weighted maneuver time + weighted terrain cost + weighted cost-to-go</p><p>The weights express preferences. I’m not assigning NASA invented coefficients: the cited description identifies factors, rather than a universal numerical weighting for every drive.</p><p>The cost-to-go asks what remains after the maneuver. I find this useful: a good next step should also leave useful subsequent options.</p><p>The paper then describes checking ranked candidates with ACE, adding its costs, and selecting the lowest-cost validated option found within the search procedure. Only an initial part of that plan is executed before replanning [5].</p><h3>A promising route still has to pass the safety check</h3><p><a href="https://arxiv.org/pdf/1808.00031">Approximate Clearance Evaluation, or ACE</a>, developed by Kyohei Otsu and colleagues, estimates minimum and maximum wheel heights over the terrain. From these it calculates conservative bounds on rover tilt, suspension angles, and body-to-ground clearance, avoiding a much more expensive iterative settling calculation [6].</p><p>That explains why the rover’s shape matters. A rock between wheels is a different geometric problem from a rock underneath a wheel. A flat footprint cannot express every interaction between a six-wheeled vehicle and uneven ground.</p><p>ACE’s conservatism is intentional: its bounds are designed to cover adverse configurations under its assumptions. That is a model-based guarantee, not proof against every possible real-world failure. A geometric clearance calculation cannot establish that all sensing, terrain, and execution assumptions are correct [6].</p><p>This is the distinction I want to keep clear: <strong>a preference can make an acceptable route more attractive; it cannot make a violated safety limit acceptable.</strong></p><h3>Route A versus Route B: make the distinction concrete</h3><p>The following is my teaching example. Its numbers are invented, and its battery budget is not ENav’s documented cost function.</p><p>Suppose a robot has two possible routes:</p><ul><li><strong>Route A:</strong> a short route that violates a tilt limit.</li><li><strong>Route B:</strong> a longer route that passes the assumed safety checks.</li></ul><p>For Route B, suppose travel needs 8 energy units, the task needs 10, and a required reserve is 5. Its total requirement is <strong>8 + 10 + 5 = 23 units</strong>.</p><p>With 30 available units, Route B fits that simplified budget. With 20, it does not. The mission must change: perhaps postpone the task or stop for new instructions, depending on the robot’s rules. Route A’s tilt violation remains a violation.</p><p>We can also illustrate weighting. Among two routes that both pass the checks, imagine normalized time and terrain costs of (4, 6) and (7, 2). Equal weights give totals of 10 and 9, favoring the second route. Doubling the time weight gives 14 and 16, favoring the first. These are arbitrary teaching scores; they demonstrate how preferences change a ranking without changing a safety boundary.</p><p>Figure 2. Original teaching diagram. All values are illustrative; this is not an AutoNav output or NASA energy model.</p><h3>Feedback: when the next move is to stop</h3><p>A student who read my earlier draft asked: <em>How does an autonomous robot decide when it should stop trusting its own plan and ask a human for help?</em></p><p>I like this question because it challenges the assumption that successful autonomy always means continued movement.</p><p>The engineering description distinguishes continued searching from failure: if ENav cannot produce a satisfactory motion in time, it can keep evaluating; if none is found, the drive fails and human operators determine the way forward. It also accounts for motion uncertainty by expanding wheel-footprint regions, including for predicted slip and more distant positions [4].</p><p>So I would separate three outcomes: reject this candidate, spend longer looking for another, or end the drive for human review. They are different decisions. There is no need to imagine the rover experiencing doubt, or a universal “confidence below 70%” rule.</p><p>My interpretation is that a useful handoff should explain the obstacle to progress. Did all considered routes fail a check? Did planning run out of time? Is better information needed? Those explanations suggest different responses. They are also questions I would ask when evaluating a rescue robot, where waiting has a different cost.</p><h3>The bottleneck: checking enough possibilities quickly</h3><p>More options are useful only if the planner can assess them in time. A deeper search can discover possibilities, but it also demands more computation and asks the system to predict farther ahead.</p><p><a href="https://arxiv.org/pdf/2203.04563">MLNav, a research framework by Shreyansh Daftry and colleagues</a>, targets this bottleneck. A learned heuristic predicts promising terrain so the planner can prioritize candidates; the model-based checker still validates selected paths. The paper reports approximately tenfold fewer collision checks in simulations reconstructed from real Martian terrain. Its hardware benchmarks measured about 12 milliseconds per ACE check on a RAD750 processor, while neural inference was benchmarked separately on a Jetson TX2. These are simulation and hardware-benchmark results, not evidence that MLNav performed those drives on Mars [7].</p><p>A faster ranking can also be wrong. In my view, the relevant question is whether its mistakes waste time, hide useful options, or allow unsafe motion. Those consequences must be measured separately.</p><p>Keeping the checker addresses one part of that question. It does not eliminate dependence on the map, the vehicle model, or the candidates the search actually considers. I would therefore want to see performance across difficult cases, including situations where stopping is the correct result.</p><h3>What real autonomy has already achieved</h3><p>There is flight evidence for AutoNav itself. In 2023, Perseverance traveled <strong>759 meters</strong> through Snowdrift Peak; the corresponding straight-line route, including two science stops, was 520 meters. JPL reported six autonomous drive sols after excluding days devoted to science. Much of the additional distance came from navigating around rocks unseen in orbital imagery [3].</p><p>For me, that is a more revealing outcome than a claim that the rover chose the shortest path. Its detours were part of reaching the destination through the terrain it actually encountered.</p><h3>My take: let learning guide the search, then demand verification</h3><p>I would support pursuing learned route-ranking methods for future rover missions, with an independently evaluated safety checker and explicit stopping conditions. The MLNav evidence makes this architecture worth investigating; it does not settle whether a particular implementation is ready to fly.</p><p>The strongest counterargument is that adding learning introduces another component to validate. A heuristic that works on familiar terrain might rank unfamiliar terrain poorly. If it repeatedly directs computation toward bad candidates, the rover could lose the very driving time the upgrade was meant to save.</p><p>What would change my mind? Repeated failures on unseen terrain, or evidence that sensing errors defeat the combined system’s safeguards, would make me more cautious. Consistent gains in integrated rover tests — with difficult terrain, realistic sensing errors, and documented safe failures — would strengthen my support. I would want the comparison to include the full computing cost, not just the checks that were avoided.</p><p>I began with a robot facing a blocked corridor. After looking at Perseverance, the question I would ask has become more precise: <strong>what evidence permits the next move, and what condition requires it to stop?</strong> That is the kind of autonomy I want to understand and help build.</p><h3>Sources</h3><ol><li><a href="https://science.nasa.gov/blog/speaking-percys-language-how-to-talk-so-perseverance-will-listen/">NASA: Speaking Percy’s Language: How to Talk so Perseverance Will Listen</a> (2021). Communication delay.</li><li><a href="https://www.nasa.gov/solar-system/nasas-self-driving-perseverance-mars-rover-takes-the-wheel/">NASA/JPL: NASA’s Self-Driving Perseverance Mars Rover ‘Takes the Wheel’</a> (2021). AutoNav and human route planning.</li><li><a href="https://www.jpl.nasa.gov/news/autonomous-systems-help-nasas-perseverance-do-more-science-on-mars/">NASA/JPL: Autonomous Systems Help NASA’s Perseverance Do More Science on Mars</a> (2023). Snowdrift Peak and Figure 1.</li><li><a href="https://www-robotics.jpl.nasa.gov/media/documents/AAS_2020_mobility_mmc_v10.pdf">McHenry et al.: Mars 2020 Autonomous Rover Navigation</a> (2020), especially pp. 10–12. Architecture, motion uncertainty, and unsuccessful drives.</li><li><a href="https://arxiv.org/pdf/2011.06022">Abcouwer et al.: Machine Learning Based Path Planning for Improved Rover Navigation</a> (2020 preprint), Section 2. Baseline ranking, experimental path tree, and execution cycle.</li><li><a href="https://arxiv.org/pdf/1808.00031">Otsu et al.: Fast Approximate Clearance Evaluation for Rovers with Articulated Suspension Systems</a> (2019 revised preprint). ACE and conservative bounds.</li><li><a href="https://arxiv.org/pdf/2203.04563">Daftry et al.: MLNav: Learning to Safely Navigate on Martian Terrains</a> (2022), Sections V–VI. Learned heuristics, hardware benchmarks, and simulation results.</li></ol><h3>About me</h3><p>I’m Aranza Pacheco, a student exploring AI, robotics, and how autonomous systems make decisions. I’m especially interested in the point where a promising idea has to become an action in the physical world — and how we decide whether that action is worth taking.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9cac13451ecd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Next Move: How Autonomous Robots Decide]]></title>
            <link>https://medium.com/@aranzapachecosuarez/the-next-move-how-autonomous-robots-decide-55a932e5f042?source=rss-8eea93e01a18------2</link>
            <guid isPermaLink="false">https://medium.com/p/55a932e5f042</guid>
            <category><![CDATA[roboticprocessautomation]]></category>
            <category><![CDATA[robotics-training]]></category>
            <category><![CDATA[robotics-automation]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Aranzapachecosuarez]]></dc:creator>
            <pubDate>Sun, 04 Oct 2026 23:25:45 GMT</pubDate>
            <atom:updated>2026-10-08T21:29:30.035Z</atom:updated>
            <content:encoded><![CDATA[<h3>From hazardous terrain to farm fields, useful autonomy depends on how machines interpret uncertainty — and how humans define the mission.</h3><p>Imagine a robot entering a damaged building. A corridor ahead is blocked. Another route is open, but exploring it will consume battery power, and the connection to its operator is weakening. Should the robot continue, return, or wait?</p><p>This is an illustrative scenario, but the planning problem is real: a useful machine must choose actions with incomplete information. Reaching a destination is only part of the challenge. It must also preserve enough resources to complete its job and respect the limits placed on it.</p><p>Autonomous robots approach such choices through a repeating process: perceive the situation, reason about possible actions, act, and use feedback to decide again. AI can strengthen this process by recognizing objects or helping evaluate options. Yet autonomy also depends on planning algorithms, control systems, and explicit rules.</p><p>The strategic question is bigger than whether a robot can move by itself: <strong>how do its individual decisions add up to a successful mission?</strong></p><h3>A mission begins before the robot moves</h3><p>Remote control gives a person direct command over a machine. Automation carries out specified processes. Autonomy allows a system to select some actions from observations and goals. These capabilities can coexist: a robot may navigate independently while a human assigns tasks or approves consequential actions.</p><p>The important question is therefore which decisions have been delegated, under what conditions, and with what fallback. A robot can be autonomous at avoiding obstacles while still depending on people to decide where it should go and why.</p><h3>Perceive, reason, act — then reconsider</h3><p><strong>PERCEIVE → REASON → ACT → FEEDBACK → PERCEIVE</strong></p><p>Figure 1. A simplified decision cycle. In practice, sensing, planning, and control may run concurrently and at different rates.</p><p><strong>Perceive: What is happening, and where am I?</strong> Sensors such as cameras, LiDAR, radar, and wheel encoders provide measurements. Depending on the system, learned models identify objects while geometric methods estimate distances, movement, and position. The result is an estimate of the world, with uncertainty, rather than a perfect copy of it.</p><p>In the building scenario, detecting a dark shape is not enough. The robot needs to estimate whether it blocks the route, whether the passage is wide enough, and how confident its location estimate is. Poor observations can turn a seemingly reasonable plan into a dangerous one.</p><p><strong>Reason: Which action best serves the goal?</strong> A navigation planner generates or searches for routes. A task planner may choose what to inspect next. These decisions involve different horizons: steering around debris addresses the next few moments; deciding whether to abandon an inspection concerns the remaining mission.</p><p>One way to compare feasible routes is a cost function that combines travel time, energy use, and estimated risk. If the system minimizes cost, greater risk should increase the score. But some actions should be prohibited altogether: a safety constraint can reject a route rather than merely make it less attractive.</p><h3>When the best route changes</h3><p>Consider two hypothetical routes to an inspection point. <strong>Route A takes two minutes</strong> but crosses a slope that exceeds the robot’s permitted operating limit. <strong>Route B takes four minutes</strong> and stays within that limit. The robot rejects Route A: a shorter travel time cannot override a safety constraint.</p><p>Now introduce battery level. Suppose taking Route B and completing the inspection requires 8 units of energy, returning afterward requires another 10, and the mission rules require a reserve of 5. Continuing therefore requires at least <strong>23 units of energy: 8 + 10 + 5.</strong></p><p>With <strong>30 units available</strong>, Route B fits the estimated budget. With <strong>only 20 units available</strong>, the same inspection no longer meets the return-and-reserve requirement. The robot must reconsider the task — perhaps returning along a feasible route, stopping safely, or requesting human guidance.</p><p><strong>Low battery does not make Route A safer. It changes whether the inspection should continue at all.</strong></p><p>Risk can change the decision too. If new sensor data suggests that Route B also exceeds a safety limit, it becomes unacceptable even with a full battery. If an obstacle instead adds a manageable detour, the planner must recalculate its energy budget before proceeding.</p><p>These numbers illustrate the logic; they are not measurements from a deployed robot. Real energy and risk estimates carry uncertainty. The lesson is that a route can remain physically passable while the mission becomes infeasible.</p><p><strong>Act: Execute, measure, and update.</strong> Controllers translate the selected action into commands for motors or other actuators. Sensors then reveal what actually happened. If the wheels slip or a new obstacle appears, the system may adjust movement, replan, or stop.</p><p>This feedback is adaptation during operation. It does not necessarily mean that the robot is retraining its AI. Reinforcement learning can train some policies through rewards, but it is one approach among several. A robot can combine learned components with programmed search and model-based checks.</p><h3>Where these decisions matter</h3><p><strong>Mars: navigating when an operator cannot steer in real time.</strong> NASA’s Perseverance rover uses AutoNav to build three-dimensional terrain maps, identify hazards, and plan routes around obstacles. Communication delays prevent Earth-based teams from simply driving it with a joystick. People still plan its activities and routes, while onboard navigation handles parts of the journey without a fresh instruction for every movement.</p><p>The mission-planning lesson is the division of responsibility. Humans establish scientific priorities; the rover resolves local navigation problems. Avoiding a rock helps only if the movement still advances the larger exploration objective.</p><p><strong>Underground exploration provides a research counterpart.</strong> In the DARPA Subterranean Challenge, JPL’s CoSTAR robots mapped environments and located simulated rescue targets. Its NeBula framework addresses uncertainty in sensing, motion, system health, and communications. These demonstrations show why exploration must account for unreliable observations and connections; they do not establish that autonomous rescue is routine in actual disasters.</p><p><strong>Agriculture: making a treatment decision plant by plant.</strong> John Deere’s See &amp; Spray uses cameras and machine learning to identify weeds and target spraying. Deere reported average herbicide savings of 59% across more than one million acres treated during the 2024 growing season. This is a manufacturer-reported result; it is not a universal reduction for every crop, product, or field.</p><p>This example shows how AI can make a narrow operational decision useful at scale: whether and where to spray. It also illustrates a boundary. Autonomous selection of spray targets does not, by itself, establish that the entire vehicle drives or manages the farm autonomously.</p><p>Across these applications, the same questions recur: what does the system observe, which actions can it choose, and what makes a choice useful? Their evidence differs. A deployed rover capability, a research demonstration, and a commercial spraying result should each be described on their own terms.</p><h3>Who decides what success means?</h3><p>A path planner answers how to reach a location. Mission planning asks which location matters, which task comes first, and whether the remaining resources justify continuing. Systems may distribute those decisions between people, onboard software, and a central dispatcher. Nav2’s documentation, for example, describes both dispatcher-led task assignment and more capable onboard autonomy that can include checking battery status and returning to a dock.</p><p>Return to the damaged-building scenario. Suppose the objective is simply to maximize the area explored. That objective could favor continued movement even when the robot has little power left to return with its observations. A more useful mission specification would also define resource reserves, reporting requirements, restricted areas, and conditions for stopping.</p><p>The distinction between objectives and constraints is crucial. Objectives guide preferences among acceptable actions. Constraints define what the system may do. Fallback procedures specify what happens when observations or operating conditions no longer support a reliable decision.</p><p>This makes strategic planning a concrete engineering task. Designers must decide what information is valuable, how to prioritize tasks, when to seek human assistance, and what evidence will demonstrate success. Those choices shape the robot’s behavior even when no one is directing its next move.</p><h3>What still limits autonomy?</h3><p>Unfamiliar conditions can weaken perception. Inaccurate maps or position estimates can undermine planning. Computation and power are finite, so spending longer evaluating options can reduce the resources available to act. Communications may also disappear precisely when an operator would be most helpful.</p><p>One promising response is to use learning selectively. In MLNav research, a learned heuristic helps prioritize candidate rover paths, while a model-based checker evaluates their safety. The authors tested the approach in simulations, including experiments using terrain data from Mars; the paper does not establish that MLNav was deployed aboard Perseverance.</p><p>That distinction captures a broader lesson: impressive demonstrations need clearly stated boundaries. A successful trial shows what happened under particular conditions. It does not establish that every future mission will succeed.</p><h3>The next move is a planning decision</h3><p>Autonomous robots can extend exploration, support work in hazardous environments, and make resource use more precise. Their value comes from connecting observations to actions that serve a defined purpose.</p><p>The perceive-reason-act cycle explains how a robot chooses its next move. Strategic mission planning explains why that move matters. A successful autonomous system needs both: the ability to respond to change, and a mission designed carefully enough that its responses remain useful.</p><h3>About the Author</h3><p>I’m Aranza, i’m a senior student i have been interested in robotics my hole life and recently starting digging on AI, i choose this topic so we can understand how every decision a robot makes as a cycle behind it, it doesn’t work that simple, i would love to hear what do you think about my article.</p><p>You can follow my writing and future explorations on <a href="https://medium.com/@aranzapachecosuarez">my Medium profile</a>., also you can connect with me on <a href="https://www.linkedin.com/in/aranza-pacheco-381ba340a?utm_source=share_via&amp;utm_content=profile&amp;utm_medium=member_ios">https://www.linkedin.com/in/aranza-pacheco-381ba340a utm_source=share_via&amp;utm_content=profile&amp;utm_medium=member_ios</a> or TKS too.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=55a932e5f042" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>