Transfer is a configuration problem before it is a model problem

A policy is usually discussed as though its weights contain the whole behavior. In practice, behavior emerges from a policy combined with calibration, sensor placement, action conventions, control timing, dependencies, and the physical task definition. A mismatch in any one layer can invalidate the result.

Public LeRobot reports show that even calibration files and identifiers can determine whether an otherwise valid workflow moves safely. That makes configuration provenance part of the executable artifact, not supplementary documentation.

1. Calibration is local and easy to lose

The same robot family does not guarantee the same joint offsets, ranges, direction conventions, or gripper behavior. LeRobot stores calibration by robot identifier and requires that identifier to remain consistent across calibration, recording, and evaluation.

  • Record the calibration method and tool version.
  • Identify calibration artifacts by digest without publishing device secrets.
  • Distinguish mechanical configuration from software normalization.

2. Sensors silently redefine the task

Camera location, resolution, frame rate, lens, crop, lighting, and synchronization change the observations available to a policy. A repository that says only “two RGB cameras” has not described a reproducible input contract.

3. Action and observation contracts drift

A matching tensor shape does not guarantee matching semantics. Joint order, units, normalization, coordinate frames, gripper direction, and end-effector conventions can differ while the software still runs.

4. Runtime timing is part of behavior

Control frequency, inference latency, dropped frames, and actuator communication affect the closed loop. Reporting model architecture without the runtime envelope can make a successful policy appear portable when it is not.

5. Success rates lack a shared denominator

A percentage cannot be compared without the task-reset procedure, success rule, trial count, object distribution, intervention policy, and failure handling. Video demonstrates possibility; a protocol and trial record establish repeatability.

Our current hypothesis

The first useful artifact is a machine-readable transfer record binding an immutable policy revision to a hardware profile, calibration fingerprint, observation/action contract, runtime envelope, and explicit evaluation protocol. This is a hypothesis to test, not a settled standard.