How Robots Are Learning From Their Own Broken Objects

A robot grips a chipped tool, loses a leg actuator or encounters a cracked plastic component. A conventional factory controller may stop because its model expects an intact machine and a pristine workpiece. New research asks a more ambitious question: could a robot observe damage, update its expectations and discover another safe way to act? The answer is emerging from several separate lines of work: resilient control, self-supervised learning, visual self-models and experiments with soft robots that can repair certain material injuries. None of this amounts to a machine feeling pain or improvising as broadly as a human mechanic. The genuine advance is more interesting: engineers are beginning to build systems that treat unexpected failure as information. What researchers discovered—and what they did not Failure as a source of data A robot is usually programmed and trained under a limited distribution of situations, while the real world offers bent grippers, worn joints and objects with unpredictable fractures. A damaged component can change both the forces a robot applies and the visual features it uses to recognize what it holds. The useful research question is whether the system can identify a mismatch before it causes repeated mistakes. To understand this part of the story, distinguish the observation from the interpretation. A robot is usually programmed and trained under a limited distribution of situations, while the real world offers bent grippers, worn joints and objects with unpredictable fractures. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. The useful research question is whether the system can identify a mismatch before it causes repeated mistakes. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Learning a model of itself A 2025 npj Robotics study by Hu, Chen and Lipson described self-supervised visual self-modeling for a robot with twelve degrees of freedom using a first-person camera. The study supports a route to predicting dynamics and responding to changed conditions, not a claim that every robot can already repair itself. A learned model can provide a reference against which unusual motions and possible damage can be assessed. To understand this part of the story, distinguish the observation from the interpretation. A 2025 npj Robotics study by Hu, Chen and Lipson described self-supervised visual self-modeling for a robot with twelve degrees of freedom using a first-person camera. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. A learned model can provide a reference against which unusual motions and possible damage can be assessed. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. The broken-object puzzle An object can fail in a thousand different ways: a ceramic cup chips, a flexible pouch tears or a wooden handle cracks along its grain. A robot trained only on the appearance of new items may misjudge the center of mass or the location of a reliable gripping surface. Resilience requires representations of functional properties, not just category labels such as cup or box. To understand this part of the story, distinguish the observation from the interpretation. An object can fail in a thousand different ways: a ceramic cup chips, a flexible pouch tears or a wooden handle cracks along its grain. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Resilience requires representations of functional properties, not just category labels such as cup or box. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Sensing what changed Cameras can reveal visible cracks while torque, force and tactile sensors detect contact changes that are impossible to infer reliably from an image alone. A robot picking up a damaged mug might sense that its handle bends under load even when the surface appears intact. Combining modalities could improve reliability but also raises calibration, cost and interpretation challenges. To understand this part of the story, distinguish the observation from the interpretation. Cameras can reveal visible cracks while torque, force and tactile sensors detect contact changes that are impossible to infer reliably from an image alone. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Combining modalities could improve reliability but also raises calibration, cost and interpretation challenges. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. The prediction error that matters A robot can compare predicted motion with observed motion and update a model when the two differ consistently. If a wheel no longer produces expected forward motion, the residual may indicate slipping, mechanical damage or a surprising surface. A discrepancy is a clue to investigate, not automatically a diagnosis of one precise broken part. To understand this part of the story, distinguish the observation from the interpretation. A robot can compare predicted motion with observed motion and update a model when the two differ consistently. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. A discrepancy is a clue to investigate, not automatically a diagnosis of one precise broken part. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Learning without a perfect answer key Self-supervised systems can train on relationships between actions and observed outcomes, reducing dependence on manually labeled examples of every possible fault. A robot may learn how the camera image should change after a careful test movement. Such learning still depends on safe data collection and may generalize badly outside the conditions represented during training. To understand this part of the story, distinguish the observation from the interpretation. Self-supervised systems can train on relationships between actions and observed outcomes, reducing dependence on manually labeled examples of every possible fault. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Such learning still depends on safe data collection and may generalize badly outside the conditions represented during training. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. A robot that adapts rather than heals Damage-adaptive control can help a machine find a new gait or manipulation strategy without repairing its hardware. A walking platform might redistribute motion among functioning limbs after one actuator becomes unreliable. Functional recovery and physical restoration are distinct achievements; an adapted gait may be slower or unsafe on rough terrain. To understand this part of the story, distinguish the observation from the interpretation. Damage-adaptive control can help a machine find a new gait or manipulation strategy without repairing its hardware. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Functional recovery and physical restoration are distinct achievements; an adapted gait may be slower or unsafe on rough terrain. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. What self-repair really means A 2024 IEEE Robotics and Automation Letters paper demonstrated a soft crawling robot using McKibben muscles and self-healing materials to reconnect particular damaged surfaces. The experiments address specific material cuts and crawling motions under controlled conditions. A laboratory self-repair mechanism does not imply autonomous replacement of circuitry, batteries or arbitrary structural components. To understand this part of the story, distinguish the observation from the interpretation. A 2024 IEEE Robotics and Automation Letters paper demonstrated a soft crawling robot using McKibben muscles and self-healing materials to reconnect particular damaged surfaces. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. A laboratory self-repair mechanism does not imply autonomous replacement of circuitry, batteries or arbitrary structural components. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. The difficulty of fragile objects Picking up broken glass or a cracked medical device introduces sharp edges, unpredictable friction and risks to nearby people. A robot might grasp the largest rigid segment rather than the object’s former handle. Safety rules may require it to refuse manipulation even when a clever controller can propose a possible grip. To understand this part of the story, distinguish the observation from the interpretation. Picking up broken glass or a cracked medical device introduces sharp edges, unpredictable friction and risks to nearby people. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Safety rules may require it to refuse manipulation even when a clever controller can propose a possible grip. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. The role of simulation Engineers can simulate missing joints, warped objects and reduced traction to generate candidate recovery policies without repeatedly breaking physical machines. A robot may practice recovering from damaged wheels inside a digital environment before trials on hardware. Simulations simplify fracture, fluid contact and sensor noise; successful virtual recovery therefore needs real-world validation. To understand this part of the story, distinguish the observation from the interpretation. Engineers can simulate missing joints, warped objects and reduced traction to generate candidate recovery policies without repeatedly breaking physical machines. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Simulations simplify fracture, fluid contact and sensor noise; successful virtual recovery therefore needs real-world validation. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Why reinforcement learning attracts attention Reinforcement learning lets an agent evaluate actions by their consequences, potentially discovering strategies that a programmer did not explicitly write. A simple objective might reward moving a damaged payload to a safe place without exceeding measured force limits. Poorly chosen reward functions can encourage unsafe shortcuts, making constraints and evaluation critical. To understand this part of the story, distinguish the observation from the interpretation. Reinforcement learning lets an agent evaluate actions by their consequences, potentially discovering strategies that a programmer did not explicitly write. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Poorly chosen reward functions can encourage unsafe shortcuts, making constraints and evaluation critical. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Safe exploration is a contradiction to manage Learning often requires trying actions whose outcomes are uncertain, yet a damaged robot should not experiment recklessly around people. A machine could perform low-force diagnostic movements in a protected zone before considering a larger action. Researchers must distinguish adaptation in a laboratory cage from deployment in a home, hospital or warehouse. To understand this part of the story, distinguish the observation from the interpretation. Learning often requires trying actions whose outcomes are uncertain, yet a damaged robot should not experiment recklessly around people. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Researchers must distinguish adaptation in a laboratory cage from deployment in a home, hospital or warehouse. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Recovering a task is not recovering trust Even after an adapted robot completes several successful trials, operators need to know whether its remaining hardware has enough margin for continued work. A fractured gripper might move a light carton successfully but fail catastrophically with a heavy item. Verification should address repeatability, load limits and failure detection rather than celebrating one impressive video. To understand this part of the story, distinguish the observation from the interpretation. Even after an adapted robot completes several successful trials, operators need to know whether its remaining hardware has enough margin for continued work. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Verification should address repeatability, load limits and failure detection rather than celebrating one impressive video. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. The problem of measuring improvement A credible benchmark should specify damage severity, test objects, task success, collateral risk, adaptation time and how often attempts fail. A system that completes seventy percent of trials after a fault can outperform a baseline and still be unsuitable for critical applications. Publishing failure cases makes it easier to see where a new algorithm has actually moved the frontier. To understand this part of the story, distinguish the observation from the interpretation. A credible benchmark should specify damage severity, test objects, task success, collateral risk, adaptation time and how often attempts fail. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Publishing failure cases makes it easier to see where a new algorithm has actually moved the frontier. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Human technicians remain essential A robot that identifies likely damage may help a technician prioritize maintenance even if the robot cannot perform repair itself. A clear log of unexpectedly high motor current and abnormal movement could narrow an inspection. Explainable alerts and reliable shutdown behaviour may create more value than pretending to eliminate skilled maintenance. To understand this part of the story, distinguish the observation from the interpretation. A robot that identifies likely damage may help a technician prioritize maintenance even if the robot cannot perform repair itself. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Explainable alerts and reliable shutdown behaviour may create more value than pretending to eliminate skilled maintenance. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. What comes next Promising research will combine visual self-modeling, tactile feedback, safe diagnostic actions and explicit damage-aware planning. A future warehouse assistant might stop, scan a warped gripper, request a replacement component and resume only after passing a test. The near-term breakthrough may be honest self-assessment and graceful degradation, not miraculous self-healing. To understand this part of the story, distinguish the observation from the interpretation. Promising research will combine visual self-modeling, tactile feedback, safe diagnostic actions and explicit damage-aware planning. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. The near-term breakthrough may be honest self-assessment and graceful degradation, not miraculous self-healing. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Why broken objects matter beyond robotics Real environments are full of repaired appliances, worn parts, reused packaging and objects whose shapes have changed since manufacture. Learning functional affordances under damage could help robots participate in recycling, inspection and maintenance. Machines that handle imperfect reality will be useful precisely because they do not assume the world stays showroom-new. To understand this part of the story, distinguish the observation from the interpretation. Real environments are full of repaired appliances, worn parts, reused packaging and objects whose shapes have changed since manufacture. That describes one piece of the mechanism; it does not by itself settle how broadly the result applies. Machines that handle imperfect reality will be useful precisely because they do not assume the world stays showroom-new. A persuasive test would specify what should happen under an alternative explanation, collect comparable observations, and report the misses as openly as the successes. Limitations, open questions, and the next experiment No single study settles all the questions raised by this topic. A useful advance must survive comparison with alternative explanations, independent replication, and testing under conditions beyond those originally selected by the researchers. The relevant uncertainties are not reasons to dismiss the research; they tell us what claims can responsibly be made today and what stronger evidence would look like tomorrow. Sources and further reading: https://doi.org/10.1038/s44182-025-00031-6; https://ieeexplore.ieee.org/document/10494870