Technical methods · Revised 12 November 2026

Humanoid robot reliability: how to draw a curve from real failures

A walking record, a task success percentage and a survival curve answer three different questions. Use the session-level tool below to see what the available data can support and which claims still need evidence.

What the AGIBOT walking claim tells us

AGIBOT's A2 Ultra product page reports up to 1,300 hours of continuous walking without anomalies. This is a manufacturer statement about an endurance test, not a published distribution of time to failure.

PublishedNot disclosed in this claimConsequence
Upper-bound walking duration of 1,300 hNumber of individual robots and sessionsCannot estimate fleet survival or the failure distribution
Manufacturer says no anomaliesDefinition of anomaly, human resets and stopping rulesCannot decide which events qualify as failures
Walking validationComplete task mix, robot configuration and environment exposureCannot infer grasp, shift-level or workplace reliability

The 1,300-hour figure is intentionally excluded from the curve below. Treating it as a failure or a right-censored session would assign details the source does not provide.

Build a reliability curve from session logs

Enter one row per independent run. The first value is the number of hours until a defined failure or the end of observation. Use failure for a recorded failure and censored when observation ends while the robot is still functioning. Keep robot model, task, failure rule and software configuration consistent across the runs you compare.

The values below are FICTIONAL teaching data. They are not measurements from AGIBOT or any other manufacturer.

10 observed runs

4 defined failures

202 total observed hours

Fictional data · illustrative curve only

0%25%50%75%100%07.51522.5301 censored at 12 hours1 censored at 17 hours1 censored at 22 hours1 censored at 24 hours1 censored at 28 hours1 censored at 30 hoursElapsed run hours

Blue steps = observed failures. Amber marks = runs that ended without the defined failure. The curve is descriptive for the entered sessions, not a forecast of product life.

Check the numbers behind the steps

HourAt riskFailuresCensoredSurvival estimate
10101090%
1290190%
1481078.8%
1770178.8%
2061065.6%
2250165.6%
2440165.6%
2531043.8%
2820143.8%
3010143.8%

Method: Kaplan-Meier product limit. At each failure time, the estimated survival fraction is multiplied by (robots at risk − failures) / robots at risk. Censored runs leave the risk set after their recorded stopping time. With no failures, the line remains at 100% and does not establish a 100% true reliability rate. Statistical assumptions include independent runs and censoring unrelated to imminent failure.

What does 27 successful tasks out of 30 mean?

A time-to-failure curve cannot answer whether a humanoid reliably picks a part, places it and returns to a ready state. For that, count complete task attempts using one agreed success criterion. Edit this separate example to calculate the observed completion rate and a 95% Wilson confidence interval.

90% observed completion rate

95% Wilson interval: 74.4% to 96.5%

The interval describes sampling uncertainty under independent, identically defined attempts. Successive runs from one robot may be correlated; a change in software, payload or lighting requires separate reporting. An interval does not remove systematic bias in the test setup.

Worked check: even if all 30 attempts succeed, a two-sided 95% Wilson interval has a lower bound near 88.6%. A short perfect run does not establish near-perfect reliability over weeks.

A field protocol that produces usable evidence

  1. Fix the actual job. For a tote-transfer test, state tote mass, pickup and drop zones, expected cycle, floor and permitted human assistance. One successful move is one attempt only.
  2. Record the hardware and controller. Track robot serial ID, hand model, software version, battery, payload and operating mode. Separate teleoperated runs from autonomous runs.
  3. Define a failure before testing. A fall, missed grasp, emergency stop, manual reset or remote recovery might end a run. Decide the rule in advance and use it for every run.
  4. Mark the reason each run stopped. Record elapsed operating hours and either failure or censored. Stopping a still-working robot at a planned end time is censoring; a broken joint is a failure.
  5. Keep the denominator. Record every attempt, failed attempt, restart, operator intervention, planned charging interval and repair duration. Do not merge scheduled charging with unplanned downtime.
  6. Publish failures alongside successes. Include trial count, start and stop conditions, intervention log, variance in task duration and a link to the protocol before comparing different robot platforms.

The minimum data log

Use these columns in a spreadsheet or CSV. One row should describe one observation with an unambiguous ending.

unit_id,run_id,robot_model,software_version,task,environment,elapsed_hours,status,stop_reason,operator_interventions,repair_minutes,task_attempts,task_successes

Metrics that should never be conflated

MetricHow to calculate or observe itLimit
Runtime or enduranceDuration walked or powered onA time interval by itself does not identify how often a defined failure occurs.
Failure-free probabilityShare of runs without a defined failure by time tNeeds individual event/censor times and a stable failure definition.
Task successSuccessful task attempts / all attemptsRequires a task boundary, a denominator and consistent treatment of human assistance.
Observed availability(scheduled minutes - unplanned downtime) / scheduled minutesDeclare treatment of planned charging, scheduled maintenance and idle periods.
Mean time between failuresOperating hours divided by counted corrective failuresOnly descriptive under a consistent failure threshold and exposure definition; no estimate from zero events.

For example, in a scheduled eight-hour shift with 45 minutes lost to unplanned faults, observed availability is 435 / 480 = 90.6%. This fictional calculation says nothing about the probability of surviving the next shift and does not disclose whether a completed task met its quality requirements.

When a reliability curve is misleading

A factory test on a flat floor can hide failures triggered by uneven surfaces or people entering the workspace. Software updates can change failure rates halfway through a trial. A robot that pauses for remote assistance may appear failure-free if the operator excludes interventions. And a censored run caused by suspicious vibration may bias the curve if the impending fault is the reason observation stopped.

The estimator is useful for comparable logged runs. It is unsuitable for inferring product lifetime from a marketing duration, aggregating incompatible tasks or assigning survival probabilities to a robot model when the number and outcome of test runs are unavailable. If some units are repeatedly repaired and returned to service, analyze recurrence separately instead of treating each repair as a new independent robot.

Sources and calculation method

The company duration is attributed to its manufacturer. The curve calculator uses the NIST Kaplan-Meier method and only the data entered in the browser. The initial sample is fictional; it is not a new observation about AGIBOT, Figure, Unitree or any other robot.

  • AGIBOT A2 Ultra – official product page

    Manufacturer walking-duration claim; test sessions and stoppage log are not provided.

  • NIST – Kaplan-Meier reliability estimates

    Non-parametric estimation from observed failure times and censored runs.

  • NIST – robot performance standards

    Repeat a defined task; state what counts as success and failure.

  • Epoch AI – Where autonomy works (2026)

    Reviews reliability, speed and ability to generalize by real task.