Files

36 lines
2.3 KiB
Markdown

# Booster T1 package
Imported from the local `mjlab` T1 flat-velocity task and `booster_assets/robots/T1/T1_23dof.xml`.
## Policy snapshot
- Training run: `2026-11-02_15-03-39_t1_flat`.
- Latest checkpoint at import: `model_6700.pt`; corresponding run-name ONNX export is bundled as `policies/walk.onnx`.
- Training may continue independently. This package intentionally pins the tested snapshot; it does not hot-reload weights.
- File hashes and source paths: `source-manifest.json`.
## Fidelity
23 controlled joints; 81 observations in exported policy order: body-frame linear velocity, angular velocity, projected gravity, relative joint positions, joint velocities, previous raw actions, and body-frame velocity command. The ONNX contains the exported observation normalization.
The MJCF is generated through mjlab's `Entity(get_t1_robot_cfg())`, preserving inertias, armatures, collision masks and native position actuators with per-joint gains and torque limits. Actions use the full-precision training scales (not the rounded ONNX metadata). Simulation timestep is 5 ms with four substeps (50 Hz policy), implicitfast/Newton. Training randomization is disabled; the platform supplies its shared floor, walls and props. Meshes are the source STL visuals; no texture replacement is included.
W/S command +0.6/-0.4 m/s; A/D command +0.6/-0.6 rad/s. Space resets, Escape pauses, C restores the chase camera. A fall holds targets pending manual reset; no recovery policy is supplied.
Browser verification: five separate 5-second standing/forward/backward/left/right trials remained upright with finite states. Forward/backward displacement was +2.84/-1.68 m; left/right yaw was +2.97/-2.82 rad. Keyboard press/release, reset and robot switching also passed. This is simulation validation, not hardware qualification.
## Re-import
From repository root:
```sh
/path/to/mjlab/.venv/bin/python app/scripts/import-t1.py /path/to/mjlab
cd app
npm run validate:packages
npx playwright test t1.spec.js
```
The importer selects the newest run and its highest numbered checkpoint, and requires a non-stale ONNX export. `BOOSTER_ASSETS_ROOT` can override the asset checkout. Re-run the browser tests after updating weights.
Licenses from both source repositories are preserved in `LICENSES/`; the policy was trained locally. Confirm asset/model redistribution rights before public deployment.