AUTON

Robot fleets fail in simulation here, and every failure comes back labeled.

AUTON generates simulated data of multi-robot fleets failing and recovering. It is for teams training and testing fleet software and robot world models.

Simulated in NVIDIA Isaac Sim 6.2 on PhysX bodies. No physical robot or real site is shown.

6,000simulated shifts run so far
8,998faults injected
330,915fleet decisions logged
8kinds of fault

Every shift answers three questions.

A simulated shift is only useful as data if you know what went wrong in it. Each run AUTON returns carries a label that says so in plain terms, and in a form a training pipeline can read.

What failed?

The fault that was injected, when, and what the robot was doing at that moment. A robot that died mid-haul is labeled differently from one that died parked.

Haul1 died at 130 s on its way to T5.

Which decision?

The fleet software decision that noticed it, the one that answered it, and, when the software caused the failure itself, the decision at the root. Each one is numbered and logged with the options it had.

Caught 6.0 s later from missed heartbeats. Its task went to Haul2.

Did the fleet recover?

Whether the displaced work was finished, how long that took, which robots were lost along the way, and how much later the shift ended than the same shift with no fault.

Yes. 16 of 16 tasks finished, 41 s later than the fault-free shift.

You send three things. Labeled shifts come back.

There is no integration to build before you can see data from your own site.

  1. Describe your operation

    The robot types you run, a map of the site, and the list of tasks a shift has to get through.

  2. We run thousands of shifts and break things

    Each shift gets one to three faults at seeded times: the kind that stop real fleets.

    • Robot dies mid-task
    • Robot immobilized
    • Degraded drive
    • Tool failure
    • Intermittent dropout
    • Dropped radio link
    • Blocked path
    • Head-on in a lane
  3. Every shift comes back labeled

    Use the set to train a coordination policy or a world model, or run it as a test suite before a release.

    yard-radio-drop

    Radio link dropped for 40 s

    Robot presumed down after 2.5 s

    T4 done twice, 41 s wasted

Replay a recorded shift.

These are real runs from the generator, drawn as a site plan. Pick one, scrub through it, and read the label it produced. Yellow marks what was injected. The fleet software never sees the yellow.

north gatesouth gate1 mT0T1T2T3T4T5T6T7T8T9T10T11T12T13T14T15EExc1EExc2HHaul1HHaul2SSurv1siteconstruction_yardpolicybaseline, seed 1

Loading this run…

0.0 of 0 s

  • fault window, as injected
  • coordinator detects it
  • displaced work finished
  • task done

Eight faults, and the fleet is never told which one.

A fault changes a robot, the radio link or the site. The fleet software only sees symptoms: missing heartbeats, a robot that says it is driving but is not moving, an obstacle someone drove up to. That gap between what happened and what was known is what the labels measure.

FaultWhat is changedWhat the coordinator can seeInjectedDetectedMedian time to detectFleet recoveredReplay
Robot dies mid-taskA robot loses power and stops where it is, for the rest of the shift.Drive actuators and the telemetry link, both cut at the same instantHeartbeats stop; the body stays in place as an obstacle1,43499%6.0 s48%
Robot immobilizedThe drive fails or the robot bogs down. It still reports, but it cannot move.Drive actuators only; telemetry keeps flowingReported position stops changing while the robot says it is driving88896%8.1 s51%
Degraded driveThe robot keeps working at a fraction of its normal speed.Drive speed limitProgress along the route falls behind the nominal rate92485%5.1 to 5.2 s84%
Tool failureOne tool stops working (bucket, bed, scanner). The robot can still drive.The tool for one named capabilityThe robot reports a tool fault after its next attempt to use that tool1,002100%2.0 s62%
Intermittent dropoutThe robot freezes and goes silent for a while, then comes back on its own.Drive actuators and telemetry together, for a fixed windowA heartbeat gap; the robot later reappears and reports again94487%6.0 s77%
Dropped radio linkThe radio link drops. The robot keeps working on its last order; the coordinator hears nothing.Telemetry and command delivery, both directionsHeartbeats stop while the robot is in fact still moving and working1,28796%6.0 s81%
Blocked pathSomething lands in a lane: a dropped pallet, a spoil heap, a parked vehicle.The site: a new rigid obstacle that is not on anyone's mapA robot senses it when it comes within range, then reports it1,276100%0.0 to 17.8 s54%
Head-on in a laneTwo robots are sent through the same single-lane aisle from opposite ends at the same time.Two relocation orders released together at opposite mouths of one aisleBoth robots report that they are blocked by the other1,243100%6.0 s74%

Percentages are over faults that had any effect on the shift. Recovered means every task the fault displaced was finished, any deadlock it caused cleared, and the shift completed. The mix includes four coordinator policies, one of them with no deadlock breaker, so these are not the scores of a tuned system. A sampled blockage always lands at the midpoint of a lane, which in the warehouse is often on a pick face; that placement rule pulls the blocked-path figure down.

Some failures are the fleet software's own doing: a deadlock from two route choices, or a task done twice because a silent robot was written off too early. Those are traced to one decision, and the shift is replayed with only that decision changed. In 4,366 such replays, the incident went away 3,676 times. The label records which.

Two ways teams use it.

As training data

For coordination policies: every decision is logged with the state it was made in, the alternatives on the table, and what happened next. For world models: rendered frames come with the ground truth for that instant, including which robot is broken and how.

As a test suite

The same seeded shifts can be run again against a new version of your fleet software. A regression shows up as specific shifts that used to recover and now do not, each with the decision that changed.

Checked against a physics engine.

The large campaigns run on simple moving bodies because that is fast. To see whether a label holds up under contact, slope and wheel slip, the same fleet software and the same fault schedule also drive articulated robots in NVIDIA Isaac Sim. The two do not always agree, and the table says where.

PhysX articulations rendered with RTX in Isaac Sim 6.2.0 on one L40S, headless. Footage plays at 12 times shift speed; the simulation itself ran at 1.9 times real time while rendering.

MeasureKinematic bodiesPhysX in Isaac Sim
Shift outcomecompletecomplete
Tasks done16/1616/16
Time to finish374.9 s381.1 s

4 of 7 schedules agree across the two backends on outcome, tasks done, detection path and recovery. Times differ because PhysX robots accelerate, turn on the spot and slip. In this run the worst body tilt was 6.1° and the closest two robots came was 0.46 m.

What this does not show.

Read these before you rely on a number from this page.

  • It is simulation. Nothing here has run on a physical robot, and no robot has been deployed to a site.
  • No customer fleet software has been connected yet. Every run uses AUTON's own baseline coordinator, which has no lane reservations and lets robots meet head-on.
  • The big campaigns have no physics. Bodies follow their routes exactly. Contact and slip are only in the Isaac Sim runs.
  • The two backends disagree on some shifts. Where a robot stands when it fails changes the outcome, and timing differs between them.
  • Faults are clean models. A dead robot stops in one step. A radio link is fully up or fully down.
  • Work is a timed pause. There is no soil model, and nothing is dug, gripped or placed physically.
  • Two sites, small fleets. Five robots in a scaled-down yard and seven full-size robots in a warehouse.
  • Sensing is simple. A robot sees an unmapped obstacle when it has line of sight within a fixed range.

Running a robot fleet, or building the software for one?

A site map, the robot types you use and a typical task list are enough to produce a first labeled set for your operation. We would like to hear what breaks for you today.

Contact Qamba Labs