Files
chenlin a8be5e0134 feat: release v0.5 with shared interactive arenas and textured RP1
Unify lighting and line-colour controls across robots, replace the football with a shared layout of draggable physical props, and import the supplied RP1 GLBs without changing trained physics or policies.

Include reproducible visual import/provenance, regression coverage and verification records. Verified 66 unit tests, 55 browser tests and production-build checks.
2026-10-30 14:14:03 +08:00

50 lines
6.6 KiB
Markdown
Raw Permalink Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
# RP1 local-evaluation package
## Inputs and provenance
- Training repository: `/home/cen/Embodied_Workspace/Mujoco_Projects/roboparty_train`.
- Model/meshes: `robolab/data/robots/roboparty/rp1`, upstream `https://github.com/Roboparty/rp1_description`, commit `d8c6dd03aeecc7f7ead3d962e09fedba4c8460f0` (clean asset worktree at original import).
- Input MJCF: `mjcf/robot_24dof.xml`; `compiler.meshdir` changes from `../` to `.` for package-relative VFS loading. XML is reserialized. Bodies, inertias, primitive collision geometry, motors, sensors and mesh bytes are retained.
- Walk: final `rp1_24dof_flat/2026-10-28_15-47-49/model_9000.pt` (9001 configured iterations).
- Dance: resumed final `rp1_24dof_beyondmimic_dance/2026-10-29_11-05-14_dance_resume/model_23998.pt` (starts from iteration 7999, adds 16000 iterations; final 23998).
- Wave: final `rp1_24dof_beyondmimic_wave/2026-10-28_19-00-23_wave/model_7999.pt` (8000 iterations).
- Importers do not launch/stop training or modify the training repository. Completed iterations do not establish convergence, arbitrary-duration stability or hardware safety.
- Logged `params/env.yaml` and `params/agent.yaml` determine actual actor branch, terms, joint order, defaults, actuator contract and network. `source-manifest.json` records checkpoint/config/motion/resource SHA-256 and export error.
## Deterministic exports
`app/scripts/import-rp1.py` exports walking, ELU MLP `243→512→256→128→24`, including empirical normalization `(obs - mean) / (std + 0.01)` and runner ±100 action clip. Three term-major history frames follow training joint order. First sample backfills history; reset/skill entry clears it. PyTorch/ONNX max absolute error: approximately **1.55e-6**.
`app/scripts/import-rp1-motion.py` exports dance/wave, ELU MLP `126→512→256→128→24`, including empirical normalization and no additional action clip (runner clip is null). Logged actor branch is **student**, not the privileged policy observation group. Inputs are reference joint positions/velocities (48), gyro (3), projected gravity (3), relative joint positions (24), joint velocities (24), last actor action (24). Each skill uses its own trained defaults and last action. The distribution and critic are not deployed. PyTorch/ONNX max errors: dance **1.31e-6**, wave **1.91e-6**.
Actions are reference-pose-relative joint targets, scale `0.25`. Named gains, torque caps and four-quadrant motor torque-speed limits are preserved. Browser intentionally omits training-time random actuator delays (0–2 physics steps), randomized physics/default positions, external pushes and observation noise. It is not an exact reproduction of Newton/MJWarp training or hardware motor control.
## Motion references and deployed limits
Source archives under `robolab/data/motions/rp1_bm/`:
- `dance1_subject2_smooth6hz.npz`: 6573 frames at 50 fps, 131.46 s total. **Only the validated first 10 s are enabled.** The previous 7999 checkpoint fell at frame 994 (~19.9 s) in a full-clip rollout. The updated 23998 checkpoint has been qualified for the 10 s prefix only; full continuous playback has not been requalified.
- `wave_R_001__A428.npz`: 218 frames at 50 fps, full 4.36 s enabled.
`.json` reference descriptors plus little-endian `.f32` files contain joint position/velocity commands in logged joint order. Motion command time advances one frame per 50 Hz control tick, freezes on pause and resets at activation. Reference end clamps; there is no random resampling or wrapping. References only supply policy observations: they do not teleport the root, set kinematic joint poses, or replace trained actions with scripted animation.
Both skills use exclusive full-body policies, a physical stop gate, entry/return blends and locomotion interruption. Validation includes three initial headings, walking before/after, repeated episodes without robot reset, keyboard/pause/reset, numerical reference checks and malformed-asset rollback. Initial integration: `docs/superpowers/verification/2026-10-29-rp1-completed-skills.md`; current dance weights: `docs/superpowers/verification/2026-10-29-rp1-dance-23998.md`. UI labels are simply `Dance` and `Wave`; durations and transition behavior remain unchanged.
## User-provided GLB material replacement
`RP1机器人glb.zip` supplies 63 Blender-exported, self-contained GLBs. `visual-source-manifest.json` records the archive SHA-256, all 63 source-part hashes, 25 output hashes and the exact body mapping. Original MJCF/STL, policy and motion files retain their existing hashes; the only manifest change is `render.body_visuals`.
- Twenty-three bodies use byte-identical source GLBs. Their Blender Y-up coordinates are rotated +90° about X into MuJoCo body Z-up. All 63 source-part bounds were independently checked against the upstream STL bounds after this axis conversion: exact agreement.
- The existing 24-DOF model has only one `left_wrist_link` / `right_wrist_link` body per side. Each visual wrist therefore composes **20 parts**: wrist roll/yaw/pitch plus the palm and 16 finger pieces. Transforms come from the copied upstream `visuals/rp1_hand_fixed.urdf`; extra wrist yaw/pitch stay neutral, fingers stay in their fixed poses. There are **no additional physical joints or finger-control claims**.
- The assembler preserves source vertex/index/image bytes, remaps GLTF indices and retains original PBR colours, roughness/metalness, normal maps, face screen texture and emissive torso logo. It adds only visual transform wrapper nodes.
- Reproduce with `python3 app/scripts/import-rp1-visuals.py /path/to/RP1机器人glb.zip app/public/robots/rp1/visuals/rp1_hand_fixed.urdf` from the repo root. All inputs/outputs validate before installation; ZIP paths are never extracted.
- Geometry additions in the hands are visual only. Existing collision envelopes intentionally remain those used by the trained 24-DOF model.
The ZIP supplied no license/attribution file. Do not infer redistribution rights from receipt of these assets. It remains a local-evaluation package under the limitations below.
## License limitation — local evaluation only
The RP1 description submodule does not supply a license file in this checkout. Do **not** infer an Apache/BSD license for its meshes/model or motion data from the platform or RoboLab. Confirm upstream rights before redistributing this package or deploying publicly.
`LICENSES/RoboLab-BSD-3-Clause.txt` preserves RoboLab's code notice; it does not grant rights to separately sourced RP1 description assets, motion recordings/retargeting data or the user's checkpoints. Training-code, model, checkpoint and motion provenance must be reviewed independently for redistribution. No public deployment was performed.