Troubleshooting¶

Missing Reach Alpha geometry¶

Eight licensed visual meshes are intentionally excluded. Obtain them through an authorized channel, review their terms and run scripts/install_reach_meshes.sh with the explicit license acknowledgment from installation. A random replacement mesh is not a valid reproduction of the frozen geometry.

Isaac Lab revision mismatch¶

The installer refuses to replace a different existing dependency checkout. Keep that environment intact and use a fresh project checkout for the pinned revision. Do not combine an editable install from one runtime with another.

First launch appears idle¶

The first RTX launch may compile shaders for several minutes. Inspect the simulator log and process activity before treating it as a failure. A short 16-step installation check can take longer to start than to execute. Avoid starting concurrent simulators while diagnosing memory or startup failures.

Artifact download reports unavailable¶

The repository is private. Use an authorized GitHub account with Git's credential helper. The downloader does not embed a token in URLs. Check the artifact ID and release access; do not paste credentials into logs. Files are verified by size and SHA-256. Different existing files are preserved, so choose a fresh destination.

Policy imports or encoder initialization fail¶

Run scripts/setup_policy_dependencies.py --verify in the intended core runtime. Use the separate HotStab dependency procedure for HotStab. Training may require pretrained encoder initialization even when inference from a full saved checkpoint does not require the raw dataset. Do not let a package upgrade replace the pinned Torch/NumPy stack while fixing an optional import.

Dataset audit or training fails¶

Check the task/schema, whole-episode success, unique seeds, camera timing, image shape and final audit selection. A pilot audit cannot authorize final training. Audit failure is evidence to inspect, not permission to relax success. See collection.

Repeated score differs¶

Check runtime revision, model hash, evaluator, resets, horizon, precision, renderer and dependencies. OpenHatch ACT has a documented 10/30 original and 16/30 instrumented repeat; numerical sensitivity is not a complete explanation. Keep the discrepancy and traces. Do not overwrite the recorded benchmark or call a tuned repeat on the same states an independent test.

Website source is missing static media¶

The Git checkout keeps code and documentation; the complete site archive includes local films, images and result JSON. Follow website restoration and build the archive as a unit. A successful build is not a public deployment.