Files
VibemakerSandbox/docs/superpowers/verification/2026-10-29-rp1-completed-skills.md

9.5 KiB
Raw Permalink Blame History

RP1 completed walking + independent dance/wave policies — 2026-10-29

Historical initial-integration record (dance checkpoint 7999). Current dance weights are checkpoint 23998; current UI labels are plain Dance / Wave. See the subsequent weight/label update for current qualification and limits.

Scope and completion evidence

User authorized an unattended integration window and subsequently requested continuation. After resuming on October 29, both BeyondMimic final checkpoints were present, stable and fully readable, with no corresponding active training process. We did not start, stop, resume or modify training jobs. Checkpoints were explicitly selected, not inferred from an arbitrary latest file.

Skill Checkpoint under roboparty_train/logs/rsl_rl Logged final iteration Export max abs error
walk rp1_24dof_flat/2026-10-28_15-47-49/model_9000.pt 9000 / 9001 configured 1.55e-6
dance rp1_24dof_beyondmimic_dance/2026-10-28_17-40-59_dance/model_7999.pt 7999 / 8000 configured 9.54e-7
wave rp1_24dof_beyondmimic_wave/2026-10-28_19-00-23_wave/model_7999.pt 7999 / 8000 configured 1.91e-6

The initial intermediate walking policy was replaced with the completed checkpoint. The selector now displays RP1, not training preview. This describes completed training iterations, not a proof of convergence or real-hardware readiness.

Actual actor contract — student, not privileged policy

The logged runner maps obs_groups.actor to [student]. Actor weights confirm [512,126] first-layer shape. The privileged observation group in the source config is not the actor input and must not be used to invent unnecessary preview/anchor inputs.

Dance/wave use float32 [1,126] → [1,24]: reference joint positions 24, reference velocities 24, body gyro 3, projected gravity 3, relative joint position 24, joint velocity 24, previous action 24. No observation history. Independent ELU MLPs include their empirical normalizers (obs - mean)/(std + .01); runner action clipping is null, and no clip is added. Each policy has its own trained default-pose map, action history and reference clock. Named joint ordering and 0.25 position-target scale are checked against logged robot/action config. PD/torque-speed parameters are checked against the existing robot contract.

Walking retains 243D three-frame term-major observations and ±100 runner clipping. All policies use 0.005 s physics / 4 decimation / 50 Hz control.

Integration and safety semantics

  • New reusable JointMotionReference reads positions/velocities in explicit float32-LE layout. It checks exact byte count, frame count, joint order, clock and finiteness. Package validation also reads/checks reference bytes and source hashes.
  • Reference data only writes observation inputs. It never writes qpos, root transforms, velocities or actuator targets; the actual trained actor generates actions.
  • Each activation resets its own reference to frame 0. Simulation time drives advancement; pause freezes it. Reference end clamps without wrapping/random resampling. Reset clears all policy state; movement interrupts to locomotion.
  • Existing exclusive FSM / stop gate / 0.5 s target blends remain shared platform behavior. Duration endpoint comparison includes a small numerical tolerance so floating simulation time cannot execute one unintended extra reference frame.
  • Missing/truncated assets fail the candidate transaction, release partial policy resources and preserve the previous robot. Retrying with good data succeeds.
  • Motion importer stages skills disabled after every update; qualification tests can explicitly enable candidates with RP1_TEST_STAGED=1. Both deployed skills were enabled only after actual physics tests passed. There are no stand-in walk policies or procedural animations in the deployed dance/wave entries.

Important negative result and deliberately bounded dance

The dance source archive is 6573 frames at 50 fps: 131.46 s. An initial full-clip browser rollout failed:

  • entered dance at simulation ~2.44 s;
  • fell at simulation 22.32 s, reference frame 994, root height 0.38572 m;
  • no non-finite values, resets or torque saturation hid the failure.

Therefore full continuous dance is not qualified or advertised. The deployed dance is explicitly the first 10 seconds, matching the training's 10 s episode horizon and original episodic skill duration. UI label is Dance / 舞蹈(10 秒片段). Its source reference remains available for future qualified extensions, but duration is 10 and deployed reference frames are 0–499. Raising duration is not a validated solution.

Wave source is 218 frames at 50 fps: the entire 4.36 s, frames 0–217, is enabled. UI label is Wave / 挥手(4.36 秒).

Positive physics qualification

app/e2e/rp1-motion.spec.js, actual MuJoCo WASM and actual exported ONNX:

  1. Each skill at three orientations reached through physical locomotion (straight, left turn, right turn), with forward walking before the request. Both complete their configured intervals, return to walking, and then make >0.4 m forward progress in 3 s without falling.
  2. Reference position/velocity observations are compared element-for-element against packaged float32 samples at the reported reference frame. Observation shape is 126, not silently padded/truncated.
  3. No non-finite state, no root teleportation: maximum per-control-step root displacement <0.014 m in the qualified trials. No sustained torque saturation (recorded saturation count zero).
  4. Minimum root height across the three wave trials approximately 0.758 m; dance approximately 0.725 m, safely above the unchanged 0.4 m fall threshold. The fall detector was not loosened.
  5. Real H/J keyboard activation; pause freezes both physics time/reference; movement interrupts and blends back; Space clears active/pending/reference/action state.
  6. wave → dance → wave → dance without resetting the robot: every episode starts at reference frame 0 and ends at its configured final frame, preserving balance.
  7. Truncated wave binary causes switch rollback, old model retention and clean retry.
  8. Recorded once-per-second upper-limb observations confirm meaningful trained motion rather than standing still: wave arm-joint tracking RMSE 0.084–0.092 rad, maximum sampled joint excursion 1.97–2.01 rad; dance RMSE 0.274–0.283 rad, excursion 1.33–1.40 rad across the three headings. These are sampled tracking metrics, not full-frame precision claims. Mid-motion production-bundle screenshots were manually inspected at reference frame 78 (app/test-results/rp1-visual-check/{wave,dance}-mid.png); wave's right wrist reached ~1.242 m while root was ~0.773 m.

app/e2e/rp1.spec.js also passes on the final walking policy: keyboard directions, falls/reset, all-family switching, three 25 s deterministic movement sequences and realtime ~50 Hz control. The walk-reused-as-episode test is still only an interface fixture, clearly separated from the real trained-motion tests.

Validation status

  • Unit tests: 53/53 passed.
  • Package validation: all four package manifests pass; RP1's 33 resource paths, both reference streams and source SHA-256 records pass.
  • Targeted RP1 browser tests: 8/8 passed on the final enabled manifest and fresh server.
  • Production build and format check passed. Existing non-fatal MuJoCo module externalization warning remains.
  • Full browser regression: 42/42 passed (10.7 minutes), including all Microduck deterministic baselines, Microban walking, all-family switching, 20 ownership/disposal transitions, race handling and deployment.
  • Production-bundle RP1 checks: 8/8 passed against a fresh Vite preview on port 5184. Reference samples, transitions, repeated activations, keyboard/reset and rollback pass outside the dev server as well. The local-only preview is left available at http://127.0.0.1:5184/; the temporary test dev server on 5183 was stopped. Existing user services were not restarted. Source/dist manifests were checked identical after the final build.
  • Final full-suite log: /tmp/rp1-completed-skills-e2e.log, exit code 0. Browser reports/screenshots are generated under ignored app/test-results/; later targeted/production runs replace the previous JSON report, while this verification record preserves the aggregate result.

An old long-running Vite server was observed returning stale transformed JS / HTML for newly added public files. Final qualification uses a separately started fresh server rather than trusting hot reload. No unrelated user server was stopped.

Reproduction and limits

npm --prefix app test
npm --prefix app run validate:packages
npm --prefix app run build
npm --prefix app run format:check
npm --prefix app run test:e2e

Use a fresh dev server; TEST_BASE_URL can select an isolated instance. Production-facing RP1 tests also run against the built bundle; development-only fixture pages are not shipped in dist.

Training-time randomized actuator delays, random physics/default poses, pushes and observation corruption remain omitted. Long dance playback, arbitrary starts, uneven terrain, severe pushes, hardware, mobile browsers and public redistribution are not accepted by these tests. RP1 description / motion redistribution rights remain unconfirmed; local evaluation only. No commit or public deployment was performed.

See RP1 operator/update documentation, source provenance and the initial unattended-window log.