LEROBOT · SO-100 · SO-101
Will this robot policy work on my hardware?
KnownRobot helps you inspect the conditions behind a published LeRobot policy before spending days integrating it. Compare the declared hardware and software contract, find missing evidence, and see measured attempts when independent teams have published them.
01
Compatibility is more than a matching robot name
A policy trained for an SO-101 can still fail on another SO-101 because calibration, camera placement, observation features, action conventions, control timing or task setup changed. KnownRobot makes those assumptions explicit without claiming that metadata alone proves physical transfer.
Run the robot-policy validator → to inventory a repository and create a portable robot-skill.yaml.
02
Six questions to answer before the robot moves
Identify the exact repository, checkpoint and framework revision. A model name or mutable branch is not a reproducible artifact.
Compare the robot family, actuator and firmware assumptions, end effector, physical modifications and safety limits.
Match camera roles, placement, resolution, preprocessing and calibration method—not merely the number of RGB inputs.
Verify feature names, joint order, units, normalization, coordinate frames and tensor shapes. Equal shapes do not prove equal meaning.
Check control frequency, inference latency, compute, drivers and dependencies that affect closed-loop behavior.
Require the reset procedure, success predicate, timeout, trials, interventions, failures and execution environment behind any reported rate.
03
What KnownRobot can tell you today
The open validator can identify declared configuration, generate a reproducibility manifest, and report missing or conflicting metadata. The local evaluator can produce measured simulator evidence. The registry connects published attempts for the same immutable policy revision so successes, failures and blocked attempts can accumulate into a compatibility history.
The registry does not yet contain enough independent evidence to promise that a particular policy will work on your arm. A clean validation result is not a physical evaluation, independent reproduction or safety certification.
04
Use KnownRobot before an integration attempt
- InspectRun the validator against the policy repository and preserve the missing-evidence report.
- CompareDescribe your robot, gripper, sensors, calibration and runtime as named features—not anonymous shapes.
- EvaluateUse a fixed task, reset and success protocol. Keep failed, blocked and unsafe-to-continue outcomes.
- PublishAttach the manifest and evidence to immutable repository revisions, then submit the attempt to KnownRobot.
START WITH YOUR REAL POLICY