<?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 Jaitly Ansh  on Medium]]></title>
        <description><![CDATA[Stories by Jaitly Ansh  on Medium]]></description>
        <link>https://medium.com/@jaitly.ansh.1?source=rss-673c90b97d5c------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*h-pI5huxtiKaxatr</url>
            <title>Stories by Jaitly Ansh  on Medium</title>
            <link>https://medium.com/@jaitly.ansh.1?source=rss-673c90b97d5c------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Wed, 18 Nov 2026 19:14:51 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@jaitly.ansh.1/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[Reliable Robots Must Know When Their Vision Is Wrong]]></title>
            <link>https://medium.com/@jaitly.ansh.1/reliable-robots-must-know-when-their-vision-is-wrong-6a68549d0ab2?source=rss-673c90b97d5c------2</link>
            <guid isPermaLink="false">https://medium.com/p/6a68549d0ab2</guid>
            <category><![CDATA[machine-learning]]></category>
            <category><![CDATA[aml]]></category>
            <category><![CDATA[robotics]]></category>
            <category><![CDATA[computer-vision]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Jaitly Ansh ]]></dc:creator>
            <pubDate>Sat, 03 Oct 2026 17:06:46 GMT</pubDate>
            <atom:updated>2026-10-24T09:15:49.857Z</atom:updated>
            <content:encoded><![CDATA[<h4>How geometry, learned estimation, and diffusion policies can connect perception to safe physical action</h4><figure><img alt="Diagram of three boxes connected by arrows: SEE (pixels and features, where are the objects), UNDERSTAND (pose, depth and memory, where am I in the world, handled by ORB-SLAM and DROID-SLAM), and ACT (possible trajectories, what should I do next, handled by Diffusion Policy). Title: A robot succeeds only when all three layers agree." src="https://cdn-images-1.medium.com/max/1024/1*mM-LDXOznZBIv8f2rDOKPA.png" /><figcaption>Seeing is only the first layer of robotic intelligence. Diagram by the author.</figcaption></figure><p>A robot can recognize a mug perfectly and still miss it by five centimeters.</p><p>The camera may label the mug correctly. The neural network may assign it 99 percent confidence. Yet when the arm moves, a small error in depth, camera pose, calibration, or timing can place the gripper beside the handle instead of around it. The vision system was right in the way an image classifier can be right. The robot still failed.</p><p>That gap between recognizing the world and acting inside it is one of the hardest problems in robotics. Working on VEX robots taught me a smaller version of the same lesson. An autonomous routine can look perfect in code, but wheel slip or an odometry error changes where the robot thinks it is. Once its internal map separates from reality, every later command can be precise and wrong.</p><p>I studied three influential papers to understand how researchers are closing that gap: ORB-SLAM, DROID-SLAM, and Diffusion Policy. They are not competing solutions to one benchmark. They solve three different layers of intelligence: building a spatial memory, improving that memory with learning, and producing actions when more than one solution may be correct.</p><p>This article asks how a vision-based manipulation robot should decide whether its spatial estimate is reliable enough to execute an action. Three distinctions help:</p><ul><li><strong>Recognition</strong>: identifying what an object is.</li><li><strong>Spatial reliability</strong>: how trustworthy its estimated location is.</li><li><strong>Action diversity</strong>: how many different movement strategies the policy can produce. Diversity alone does not measure safety.</li></ul><blockquote><em>Safe robotic action depends not only on what a machine sees, but also on whether its understanding is reliable enough to use.</em></blockquote><h3>The problem begins after recognition</h3><p>Computer vision often starts with <strong>classification</strong>, naming what is in an image; <strong>detection</strong>, locating it with a box; and <strong>segmentation</strong>, marking its pixels. A moving robot also needs the object’s position in a coordinate system that stays consistent over time.</p><p>A robot needs answers that persist across time. How did the camera move between frames? Which pixels correspond to the same physical point? Is the object two centimeters away or twenty? If the camera returns to the same room, is that a new place or a loop in the path? And if the estimate becomes unreliable, should the robot continue moving, slow down, or stop?</p><p>These questions turn vision into <strong>state estimation</strong>: updating the robot’s position and its model of the world as new observations arrive.</p><h3>ORB-SLAM gave the robot a memory</h3><p>ORB-SLAM, created by Raúl Mur-Artal, J. M. M. Montiel, and Juan D. Tardós at the University of Zaragoza, performs <strong>simultaneous localization and mapping (SLAM)</strong>: estimating a camera’s position while building a map of its surroundings [1]. It is <strong>monocular</strong>, meaning it uses one camera. The map and camera trajectory must be estimated together because neither is known in advance.</p><h4>One set of ORB features serves four jobs at once</h4><p>ORB-SLAM uses the same <strong>ORB features</strong>, distinctive image points paired with compact binary descriptions, for four jobs. <strong>Tracking</strong> estimates the current camera position and orientation, or <strong>pose</strong>. Mapping builds the 3D scene. <strong>Relocalization</strong> recovers the camera’s position after tracking fails. <strong>Loop closure</strong> recognizes a previously visited place and corrects accumulated drift.</p><p>Using one feature representation across these jobs lets the robot reuse matches instead of maintaining separate descriptions of the same scene. Local mapping estimates 3D points from multiple views and decides which frames add useful information.</p><p><strong>Keyframes</strong> are selected images kept as references for the map. ORB-SLAM retains useful views and removes redundant ones, so its memory grows with the environment rather than every second of video.</p><h4>Bundle adjustment keeps ORB-SLAM’s map consistent as the robot moves</h4><p><strong>Bundle adjustment</strong> jointly refines camera poses and 3D points. Its target is <strong>reprojection error</strong>: the pixel distance between where a mapped point is predicted to appear and where it was actually observed. Reducing that error makes the views more geometrically consistent.</p><p>The geometric error has a physical meaning: a mapped point should project back to its observed pixel. A low error is useful evidence, but it does not by itself guarantee a correct map or a safe action.</p><p>ORB-SLAM can also refuse to initialize a map when its geometric checks are ambiguous. It needs enough <strong>parallax</strong>, the apparent shift of nearby points relative to distant ones as the camera moves, to reconstruct the scene. Waiting can prevent a bad map from being created. This is a useful rejection rule, not a calibrated safety guarantee [1].</p><h3>ORB-SLAM breaks on low texture, motion blur, and repeated patterns</h3><p>Feature-based SLAM needs repeatable landmarks. Blank walls offer few corners, motion blur smears them, and repeated patterns create false matches. Moving people also violate the static-scene assumption. Monocular vision alone cannot recover <strong>metric scale</strong>, the map’s real size in meters: multiplying every distance and camera translation by the same factor leaves the images unchanged.</p><p>The system is interpretable and efficient, but its hand-designed features limit which image evidence it can use. DROID-SLAM asks whether learning can improve the difficult front end without throwing away the geometric back end.</p><h3>DROID-SLAM taught geometry to learn</h3><p>Developed by Zachary Teed and Jia Deng at Princeton, DROID-SLAM uses <strong>dense correspondence</strong>, matching locations across many image pixels rather than selected keypoints. It estimates camera poses and <strong>inverse depth</strong>, one divided by distance, across active frames. A recurrent network proposes matching corrections and weights their influence on geometric optimization [2]. These weights are not calibrated probabilities that the robot’s pose is correct.</p><p>The network repeatedly proposes corrections. A <strong>differentiable bundle-adjustment layer</strong>, an optimizer that can participate in neural-network training, turns them into pose and depth updates constrained by camera geometry. Those updated estimates feed the next round. Learning improves matching while geometry keeps the views consistent.</p><blockquote><em>DROID-SLAM is powerful because learning and geometry are inside the same feedback loop.</em></blockquote><h4>Dense correspondence lets DROID-SLAM keep tracking where sparse features fail</h4><p>ORB-SLAM matches selected keypoints. DROID-SLAM can use evidence spread across much more of the image. Its learned weights reduce the influence of unreliable matches, although they cannot guarantee recovery from moving objects or unfamiliar scenes.</p><p>The model was trained on synthetic monocular video from <strong>TartanAir</strong> [11] and tested without retraining on other camera types and datasets. On monocular <strong>EuRoC</strong> [12], mean absolute trajectory error was 2.2 centimeters ([2], Table 4). That measures camera-path error after evaluation alignment, not a robot arm’s grasp error. The ETH3D evaluation [14] also tested tracking reliability ([2], Figure 4).</p><p><strong>Stereo</strong> uses two cameras; <strong>RGB-D</strong> combines color images with measured depth. The dataset papers [11–14] describe these test conditions. Success on them does not establish reliability in every deployment.</p><p>That generalization is possible partly because the learned features do not carry the whole problem. At inference time, geometry still decides how pose, depth, and camera projection must agree.</p><h4>Learned estimation costs compute, memory, and latency</h4><p>The paper’s real-time setup uses two RTX 3090 GPUs. Longer EuRoC, TartanAir, and ETH3D sequences require a GPU with 24 GB of memory, while the TUM RGB-D results [13] can be reproduced on a single GTX 1080 Ti. With reduced resolution and frame skipping, the authors report 20 frames per second on EuRoC and 30 on TUM RGB-D, versus 8 on faster-moving TartanAir ([2], “Timing and Memory”). These are hardware-specific measurements, not universal requirements.</p><p>For a mobile robot, compute becomes part of intelligence. A more accurate estimate that arrives too late may be less useful than a slightly weaker estimate that arrives before the robot collides. Any real deployment must evaluate trajectory error, latency, memory, energy, and recovery together.</p><h3>Diffusion Policy taught robots to choose among futures</h3><p>SLAM estimates where the robot is. It does not decide what the robot should do. Cheng Chi, Shuran Song, and their collaborators developed Diffusion Policy to address that next layer by learning robot behavior from demonstrations [3].</p><p><strong>Behavior cloning</strong> learns actions from demonstrated examples. A single regression output can fail when several strategies work: averaging demonstrations that go left and right around an obstacle may point straight into it.</p><figure><img alt="Diagram of a robot on the left and a goal on the right with an obstacle between them. One curved path goes over the obstacle and another goes under it, both labeled valid. A dashed straight arrow labeled average action runs from the robot directly into the obstacle." src="https://cdn-images-1.medium.com/max/1024/1*v4gZK44_r_X99Fu1hEWcuA.png" /><figcaption>Two valid action modes do not imply that their average is valid. Diagram by the author.</figcaption></figure><h4>Diffusion denoising preserves distinct action strategies</h4><p>Diffusion Policy begins with random noise representing a future action sequence. Through <strong>diffusion denoising</strong>, repeated learned corrections turn that noise into actions conditioned on recent observations. Different initial noise samples can produce different strategies.</p><p>Predicting a sequence helps the robot stay with one strategy. It executes only the first part, observes again, and replans. This is <strong>receding-horizon control</strong>: repeatedly updating a plan while committing to only its next short segment.</p><p>The RSS paper evaluates 12 tasks across four benchmark sources [15–18] and reports a 46.9 percent mean relative improvement over the strongest baseline for each task ([3], Tables I–II and Appendix B). That is a relative change, not percentage points. <strong>DDIM sampling</strong>, a method that uses fewer denoising steps at inference, reduces 100 training steps to 10 inference steps in the paper’s timing example, reaching about 0.1 seconds on an NVIDIA RTX 3080 ([3], Section III-D).</p><h3>These papers solve different failure layers</h3><p>It is tempting to ask which of the three systems is best. That is the wrong comparison. ORB-SLAM and DROID-SLAM estimate state. Diffusion Policy generates action. A robot needs both.</p><figure><img alt="Table comparing ORB-SLAM, DROID-SLAM, and Diffusion Policy: spatial memory, role of learning, failure modes, and compute costs. ORB-SLAM uses sparse geometry; DROID-SLAM learns dense matches; Diffusion Policy generates action sequences. Their outputs address different stages and do not independently certify safe motion." src="https://cdn-images-1.medium.com/max/1024/1*r4pUrAXfEdaCyRmqdCAwcQ.png" /><figcaption>Each system solves a different failure layer. ORB-SLAM rejects inconsistent geometry; DROID-SLAM weights correspondences; Diffusion Policy samples actions. None of these signals alone certifies a safe reach. Table by the author, based on [1–3].</figcaption></figure><h3>The missing connection is uncertainty</h3><p>The most important difference among these systems is not whether they use neural networks. It is what their uncertainty means.</p><p>ORB-SLAM treats some matches as outliers and refuses to initialize when the geometry is ambiguous. DROID-SLAM predicts continuous confidence weights for dense correspondence. Diffusion Policy samples different action sequences, which can represent several valid strategies demonstrated by humans.</p><p>Those signals should not be merged into one generic confidence score. Variation among diffusion samples may be healthy: moving left and moving right could both work. Low geometric confidence is different. If the robot has the wrong target pose, every action sample may be based on the same false state.</p><p>This creates a control rule. Action diversity should influence which path the robot chooses. Geometric uncertainty should influence whether the robot moves at all.</p><h3>Robots will be safer when they report two separate confidence numbers</h3><p>My hypothesis is that manipulation robots will fail less often if they keep two reliability estimates separate: the probability that the target pose is within a task-specific tolerance, and the probability that a proposed action will succeed given that state. Low spatial reliability should block a new reach and trigger a safe pause or another observation. Several different actions can all be valid; sample diversity alone is not the second probability.</p><figure><img alt="Proposed flowchart: observe the scene and estimate target-pose reliability. If below threshold, pause and look again. Otherwise propose an action and estimate its success probability given the state. Low action confidence triggers replanning or help. Commands that pass both checks still face speed, force, and workspace limits before a short execution segment. Then observe again. Action diversity alone is not success confidence." src="https://cdn-images-1.medium.com/max/1024/1*SXW_CxChq7R63APjaj7NLA.png" /><figcaption>Proposed control rule: check spatial reliability, then action-success reliability, then enforce hard motion limits. Low confidence triggers a safe pause, another observation, or replanning. This is a hypothesis to test, not a safety guarantee. Diagram by the author.</figcaption></figure><p>I expect part of the improvement to come from estimation and control software. DROID-SLAM’s experiments compare algorithms on the same recorded images, so its gains cannot come from upgrading the camera ([2], Table 4). That supports a narrower claim than saying cameras do not matter: better sensors can help, but a clearer image alone does not tell a controller when its estimate is trustworthy.</p><p>I focus on one company example: NVIDIA’s <strong>Isaac GR00T N1</strong>, a model that turns camera observations, robot state, and language instructions into motor actions. NVIDIA’s announcement [4] describes the product; its technical report [7] supplies the architecture and evaluations. Both are company-authored evidence, not independent safety validation.</p><p>The report describes a vision-language module coupled to an action generator trained with <strong>flow matching</strong>, which learns to transform noise into actions. Its simulation Table 2 compares GR00T N1 against behavior-cloning and Diffusion Policy baselines; Section 4.2 describes real-robot manipulation tests. Those evaluations support task capability, but do not establish reliable uncertainty detection. In Sections 2.1, 4.4, and 4.6 of the March 2025 report, I did not find separately calibrated target-pose and action-success estimates linked to a stop rule. This is a limit of the documented evidence I reviewed, not proof that NVIDIA has no internal safety system.</p><p>Related work offers pieces of the problem. Kendall and Gal estimate depth uncertainty from image noise and model uncertainty [8]. KnowNo uses <strong>conformal prediction</strong>, a method calibrated on held-out examples to form prediction sets, to decide when a language-model planner should ask for help [9]. That concerns high-level choices, not continuous motor safety. These papers and GR00T N1’s report motivate my proposal; they do not establish that no existing system combines these signals.</p><p><strong>Confidence calibration</strong> means that predictions assigned a given probability are correct about that often. Guo and colleagues found miscalibration in image and document classifiers [10]. That does not directly measure robot-pose uncertainty, but it explains why a network’s confidence needs testing. I would keep speed and force limits, workspace constraints, and a human override outside the learned policy. A separate safety layer has also been studied for generalist policies [6]; its guarantees depend on its model and constraints.</p><p>Here is what could prove me wrong. In simulation, keep the camera, policy, and task fixed. Compare an ungated policy, one combined confidence score, and two separate estimates on held-out occlusions and calibration shifts. Choose thresholds on separate validation trials and compare systems at matched task completion. If the two estimates fail to predict dangerous errors, or fail to reduce collisions and false-confidence failures without simply refusing more tasks, the proposed advantage is unsupported. I have not run this experiment yet.</p><h3>Uncertainty-aware control matters wherever robots share space with people</h3><p>A few concrete examples show why this is not just an academic distinction:</p><ul><li>A warehouse robot may need to slow when a pallet hides part of its path.</li><li>An assistive robot should pause when it loses a reliable estimate of a person’s hand or a medication container.</li><li>An inspection robot should remember where evidence was collected while marking regions where localization was weak.</li></ul><p>In each case, predictable failure behavior matters more than an impressive average benchmark score.</p><p>Deployment adds other costs. Visual maps can expose scene details through reconstruction attacks [5]. Demonstration quality affects learned behavior [15], so training data should be checked for unsafe shortcuts. The GPU use reported in [2] and [3] also creates an energy budget, although those papers do not provide a deployment energy comparison. I would measure joules per completed task, minimize stored scene data, and log interventions without retaining unnecessary personal imagery.</p><h3>Conclusion: reliable robots need uncertainty-aware control</h3><p>Before reading these papers together, I thought of computer vision and robot control as mostly separate stages. First see, then act. The deeper lesson is that perception and action should continuously challenge each other.</p><p>ORB-SLAM shows why a robot needs memory and recovery. DROID-SLAM shows that learned visual features become more reliable when geometry constrains them. Diffusion Policy shows why one “best” action is often the wrong model of human behavior. Together, they show that reliable action depends on both the action plan and the quality of the state estimate.</p><blockquote><em>A capable robot does not only predict the world correctly. It changes its behavior when that prediction becomes uncertain.</em></blockquote><h3>Test robots for false confidence before trusting them</h3><p>If you build or evaluate a robot, do not report task success alone. Test what happens under occlusion, camera motion, poor texture, moving distractors, and calibration error. Report on each of the following:</p><ul><li>Confidence calibration: predicted reliability versus observed correctness, including under disturbances.</li><li>Collisions: contact events per trial, with severity recorded.</li><li>Human interventions: how often a person must take over.</li><li>Recovery time: time needed to regain a usable estimate and resume.</li><li>Latency: time from observation to a usable control decision.</li><li>False-confidence failures: incorrect or dangerous actions taken while reported reliability remains above the execution threshold.</li></ul><p>A robot that succeeds 90 percent of the time but cannot recognize the dangerous 10 percent is not ready for people to trust it.</p><p>For a first student experiment, reproduce one vision-based policy in simulation and introduce one controlled visual disturbance. Hold out disturbance settings from calibration and measure whether confidence predicts failure. Share your setup and results in a comment below this article, including cases where the confidence measure did not work.</p><h3>Sources and further reading</h3><p>[1] Mur-Artal, R., Montiel, J. M. M., and Tardós, J. D. ORB-SLAM: A Versatile and Accurate Monocular SLAM System. IEEE Transactions on Robotics, 2015. <a href="https://doi.org/10.1109/TRO.2015.2463671">https://doi.org/10.1109/TRO.2015.2463671</a></p><p>[2] Teed, Z., and Deng, J. DROID-SLAM: Deep Visual SLAM for Monocular, Stereo, and RGB-D Cameras. NeurIPS, 2021. <a href="https://proceedings.neurips.cc/paper/2021/hash/89fcd07f20b6785b92134bd6c1d0fa42-Abstract.html">Proceedings paper</a>. Version consulted for Table 4, Figure 4, and timing: <a href="https://arxiv.org/abs/2108.10869v1">https://arxiv.org/abs/2108.10869v1</a></p><p>[3] Chi, C., et al. Diffusion Policy: Visuomotor Policy Learning via Action Diffusion. RSS, 2023. <a href="https://doi.org/10.15607/RSS.2023.XIX.026">https://doi.org/10.15607/RSS.2023.XIX.026</a>. Version consulted for tables and appendices: <a href="https://arxiv.org/abs/2303.04137v4">https://arxiv.org/abs/2303.04137v4</a></p><p>[4] NVIDIA. NVIDIA Announces Isaac GR00T N1 — the World’s First Open Humanoid Robot Foundation Model — and Simulation Frameworks to Speed Robot Development. NVIDIA Newsroom, March 2025. <a href="https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks">https://nvidianews.nvidia.com/news/nvidia-isaac-gr00t-n1-open-humanoid-robot-foundation-model-simulation-frameworks</a></p><p>[5] Shibuya, M., Sumikura, S., and Sakurada, K. Privacy Preserving Visual SLAM. ECCV, 2020. <a href="https://arxiv.org/abs/2007.10361">https://arxiv.org/abs/2007.10361</a></p><p>[6] Tölle, M., et al. Towards Safe Robot Foundation Models. Preprint, 2025. <a href="https://arxiv.org/abs/2503.07404">https://arxiv.org/abs/2503.07404</a></p><p>[7] NVIDIA (Bjorck, J., et al.). GR00T N1: An Open Foundation Model for Generalist Humanoid Robots. Technical report, arXiv:2503.14734, 2025. <a href="https://arxiv.org/abs/2503.14734">https://arxiv.org/abs/2503.14734</a></p><p>[8] Kendall, A., and Gal, Y. What Uncertainties Do We Need in Bayesian Deep Learning for Computer Vision? Advances in Neural Information Processing Systems 30, 2017. <a href="https://papers.nips.cc/paper/2017/hash/2650d6089a6d640c5e85b2b88265dc2b-Abstract.html">https://papers.nips.cc/paper/2017/hash/2650d6089a6d640c5e85b2b88265dc2b-Abstract.html</a></p><p>[9] Ren, A. Z., Dixit, A., Bodrova, A., Singh, S., Tu, S., Brown, N., Xu, P., Takayama, L., Xia, F., Varley, J., Xu, Z., Sadigh, D., Zeng, A., and Majumdar, A. Robots That Ask For Help: Uncertainty Alignment for Large Language Model Planners. Proceedings of the 7th Conference on Robot Learning, PMLR 229, 2023. <a href="https://proceedings.mlr.press/v229/ren23a.html">https://proceedings.mlr.press/v229/ren23a.html</a></p><p>[10] Guo, C., Pleiss, G., Sun, Y., and Weinberger, K. Q. On Calibration of Modern Neural Networks. ICML, PMLR 70, 2017. <a href="https://proceedings.mlr.press/v70/guo17a.html">https://proceedings.mlr.press/v70/guo17a.html</a></p><p>[11] Wang, W., et al. TartanAir: A Dataset to Push the Limits of Visual SLAM. IROS, 2020. <a href="https://arxiv.org/abs/2003.14338">https://arxiv.org/abs/2003.14338</a></p><p>[12] Burri, M., et al. The EuRoC Micro Aerial Vehicle Datasets. IJRR, 2016. <a href="https://doi.org/10.1177/0278364915620033">https://doi.org/10.1177/0278364915620033</a></p><p>[13] Sturm, J., et al. A Benchmark for the Evaluation of RGB-D SLAM Systems. IROS, 2012. <a href="https://cvg.cit.tum.de/_media/spezial/bib/sturm12iros.pdf">https://cvg.cit.tum.de/_media/spezial/bib/sturm12iros.pdf</a></p><p>[14] Schöps, T., Sattler, T., and Pollefeys, M. BAD SLAM: Bundle Adjusted Direct RGB-D SLAM. CVPR, 2019. ETH3D-SLAM benchmark. <a href="https://openaccess.thecvf.com/content_CVPR_2019/html/Schops_BAD_SLAM_Bundle_Adjusted_Direct_RGB-D_SLAM_CVPR_2019_paper.html">https://openaccess.thecvf.com/content_CVPR_2019/html/Schops_BAD_SLAM_Bundle_Adjusted_Direct_RGB-D_SLAM_CVPR_2019_paper.html</a></p><p>[15] Mandlekar, A., et al. What Matters in Learning from Offline Human Demonstrations for Robot Manipulation. CoRL 2021, PMLR 164, 2022. RoboMimic. <a href="https://proceedings.mlr.press/v164/mandlekar22a.html">https://proceedings.mlr.press/v164/mandlekar22a.html</a></p><p>[16] Florence, P., et al. Implicit Behavioral Cloning. CoRL 2021, PMLR 164, 2022. Push-T and block pushing. <a href="https://proceedings.mlr.press/v164/florence22a.html">https://proceedings.mlr.press/v164/florence22a.html</a></p><p>[17] Shafiullah, N. M., et al. Behavior Transformers: Cloning k Modes with One Stone. NeurIPS, 2022. Multimodal benchmark data. <a href="https://arxiv.org/abs/2206.11251">https://arxiv.org/abs/2206.11251</a></p><p>[18] Gupta, A., et al. Relay Policy Learning: Solving Long-Horizon Tasks via Imitation and Reinforcement Learning. CoRL 2019, PMLR 100, 2020. Franka Kitchen. <a href="https://proceedings.mlr.press/v100/gupta20a.html">https://proceedings.mlr.press/v100/gupta20a.html</a></p><h3>About the author</h3><p>Ansh Jaitly is a high school junior in the Bay Area and co-captain of VEX V5 Team 3134G. He also captains VEX AI Team 3303X, which qualified for the 2025–26 VEX AI World Championship and won the Build Award there. He CAD-designs robots in Onshape, writes autonomous and odometry code in C++, and got his first hands-on computer vision experience (YOLO object detection) through the Inspirit AI Scholars Program. Outside of competition, he’s building an autonomous airport baggage robot: a wheeled robot with an arm that picks bags out of a pile, loads them into a cart, and drives them to the plane. That’s why this topic matters to him: an arm pulling a suitcase out of a messy pile has to know when its camera has lost track of the bag, and stop instead of guessing.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6a68549d0ab2" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>