| name | diagnose-holomotion |
| description | Diagnose HoloMotion errors across environment setup, motion conversion, retargeting, training, evaluation, export, simulation, offline deployment, and live teleoperation. Use when a command fails, hangs, produces invalid motion or policy output, differs between stages, or has an unknown HoloMotion root cause. |
Diagnose HoloMotion
The goal is either to solve a known failure or to localize an unknown failure to one stage with reproducible evidence.
Establish the failing boundary
Collect:
- exact command and working directory;
- relevant environment and config names without secrets;
- complete first error, not only the final wrapper error;
- input and output paths and schemas;
- last known working revision or artifact;
- expected versus observed behavior;
- whether the failure is deterministic;
- whether robot actions were involved.
Classify the failure:
- environment and dependencies;
- raw motion parsing;
- HoloSMPL canonical/formal conversion;
- HoloRetarget;
- robot HDF5 loading;
- training;
- checkpoint evaluation;
- PyTorch-to-ONNX export;
- MuJoCo sim2sim;
- Docker and robot runtime;
- offline reference;
- teleoperation reference;
- policy inference or physical response.
Do not debug later stages until their inputs are validated.
Reduce the problem
- Read the actual entry script and selected config.
- Reproduce with the smallest representative input and resource count.
- Validate shapes, dtypes, units, coordinate frames, joint order, frame rate, and metadata at the failing boundary.
- Compare a failing sample with a known-good public sample.
- Find the first divergent artifact when two backends or stages disagree.
- Search source by the exact exception, symbol, or log message.
- Change one variable at a time and preserve the command and evidence.
Use the holomotion-train-interpreter skill for training-environment Python. Use isaaclab-source-lookup for Isaac Lab implementation questions.
Safety and cost
- Prefer read-only inspection and no-action checks.
- Do not submit remote jobs or start costly training without explicit confirmation.
- Never operate the robot or bypass deployment safety checks without explicit confirmation.
- Do not ask users to publish private data, credentials, checkpoints, or machine-specific paths.
Report
Report:
- symptom;
- last successful stage;
- first failing stage;
- root cause or bounded hypotheses;
- evidence for each conclusion;
- minimal fix or next discriminating check;
- validation performed;
- remaining runtime or real-robot validation.