Robot recovery: when retrying helps, and when it hides failure

Learning goals and prerequisites

You should know conditional probabilities, basic feedback control and the distinction between an observation and an action. Learn to separate policy learning from recovery-controller changes, specify action-label semantics, and evaluate success under a fixed execution budget.

Predict first: a robot recovers twice and finishes a scripted trace. Is this evidence of task completion? What if recovery used more time than the baseline’s entire episode?

Three recent reading briefs

Connect the ideas to an actual robot stack

Consider a drawer-opening assistant. A demonstration records a handle grasp and a pull. Training on measured joint positions instead of commanded targets can silently change what the policy learns to output. At deployment, a failure predictor may notice a missed grasp, but reversing the arm does not guarantee that the drawer or object returns to its original state. A video-derived reference can also be kinematically plausible while exceeding the tracker’s limits.

A threshold is a decision rule, not a guarantee

Let p be a calibrated failure probability, C_F the loss from continuing into failure, and C_R a fixed recovery cost. In a deliberately simplified model where recovery prevents failure completely, recover when:

If C_F=100 and C_R=20 in the same utility units, the threshold is 0.2. If recovery itself fails with probability q=0.1, its expected cost becomes 20+0.1×100=30; the threshold becomes 0.3 under these assumptions. This ignores changing state, future options and risk constraints. Do not use it as a physical safety controller. The demo’s threshold 0.7 is a configurable fixture value, not an estimate from any of the papers.

A scalar risk estimate is dimensionless. Joint targets might be radians, Cartesian targets metres, and delays milliseconds; mixing these in an unweighted cost is invalid. For a horizon H and d joints, explicitly distinguish command arrays of shape H×d from observation arrays sampled at a different frequency.

Run a bounded recovery experiment locally

The synthetic risk stream [0.2, 0.8, 0.3] triggers one recovery at threshold 0.7. With one permitted recovery, [0.8, 0.9] stops at the second high-risk step. No neural network, real robot, API or upload is involved.

Expected: seven PASS checks; trace-ended is not a success verdict.

Download complete implementation. Node.js 22+, no dependencies and no randomness.

node --input-type=module -e "import {run} from './robot-recovery.mjs'; console.log(run())"

// Original teaching state machine. No robot commands, network or learned predictor.
export function simulate({risks,threshold=0.7,maxRecoveries=1,safe=true}) {
 if(!Array.isArray(risks)||!risks.length||risks.some(x=>!Number.isFinite(x)||x<0||x>1)||!Number.isFinite(threshold)||threshold<0||threshold>1||!Number.isInteger(maxRecoveries)||maxRecoveries<0||typeof safe!=='boolean')throw Error('Invalid configuration');
 let recoveries=0;const events=[];
 for(const risk of risks){
  if(risk>=threshold){
   if(!safe){events.push('STOP_UNSAFE');return {events,recoveries,status:'stopped'};}
   if(recoveries>=maxRecoveries){events.push('STOP_BUDGET');return {events,recoveries,status:'stopped'};}
   events.push('PAUSE','RECOVER','RESUME');recoveries++;
  }else events.push('ACT');
 }
 return {events,recoveries,status:'trace-ended'};
}
export function run(){const checks=[];const check=(name,ok)=>{if(!ok)throw Error(name);checks.push('PASS: '+name)};
 const recovered=simulate({risks:[.2,.8,.3]});check('one bounded recovery',recovered.recoveries===1&&recovered.events.join(',')==='ACT,PAUSE,RECOVER,RESUME,ACT');
 check('unsafe recovery stops',simulate({risks:[.8],safe:false}).events[0]==='STOP_UNSAFE');
 check('repeated high risk exhausts budget',simulate({risks:[.8,.9]}).status==='stopped');
 check('zero budget stops without recovery',simulate({risks:[.7],maxRecoveries:0}).recoveries===0);
 check('threshold equality triggers',simulate({risks:[.7]}).recoveries===1);
 let rejected=0;for(const risks of [[],[NaN],[-.1],[1.1]])try{simulate({risks})}catch{rejected++}check('invalid probabilities rejected',rejected===4);
 check('trace end is not task success',recovered.status==='trace-ended');
 return {checks,recovered,boundary:'A scripted risk stream tests control logic, not a learned failure predictor or a physical robot.'};}

The safe flag stands in for an independently validated recovery check; it does not compute collision-free motion. The script tests state transitions, threshold boundaries, invalid probabilities and budget exhaustion. Real code must also handle sensor loss, asynchronous cancellation and emergency stops.

A reproducible evaluation protocol

1. Freeze the policy checkpoint, controller versions, task set, reset distribution and terminal success evaluator.
2. Compare no recovery, random intervention and predictor-driven recovery under equal total episode time. Count pauses and recovery motions in that budget.
3. Fit and calibrate the failure predictor on development rollouts only. Split by task or scene where appropriate, not adjacent frames from the same rollout.
4. Report paired task outcomes, intervention precision/recall, unnecessary interruptions, recovery duration and final success. A detector’s accuracy alone is insufficient.
5. Test unseen objects and scenes, several thresholds and a separate tracker. Select thresholds before opening the held-out results.

For example, 60 baseline successes versus 68 recovery successes among 100 episodes is an 8 percentage-point difference. It is not enough to establish significance without the paired outcomes and sampling assumptions. A frozen policy with a better controller is a system-level improvement; it does not prove improved model weights or general recursive self-improvement.

Original interview questions and exercises

Transfer exercise: add a cooldown measured in observation steps. Test a risk stream that oscillates around 0.7. Explain why hysteresis can reduce chattering but delay a legitimate new intervention.

Continue learning

Physical AI curriculum · Reward gains versus capability gains · DPO foundations

Sources were checked on 26 October 2026. The linked paper pages provide author-maintained project links where available; no unverified code release or conference acceptance is claimed. Read the full methods before reproducing the papers.