5.3 KiB
RP1 platform integration verification — 2026-10-28
Artifact
- Package:
app/public/robots/rp1/, third family in the existing selector; generic runtime, no RP1-id branches in physics/rendering/runtime management. - Source description:
rp1_description@d8c6dd03aeecc7f7ead3d962e09fedba4c8460f0. - Checkpoint:
rp1_24dof_flat/2026-10-28_15-47-49/model_3000.pt, SHA-256ac57cfa7257b0cee309536af695f4f793be4b82bf106daa3372a0ae6da477931. - ONNX SHA-256:
0b89b7b4c1c35dfb1912a0a39950e283f31c760d8e386ad8e2da20851cffc34a. - Export includes empirical normalization and deterministic ELU actor. ONNX checker + zero/eight seeded random comparisons against PyTorch: max absolute error
9.5367431640625e-7. - Both full import and
--policy-onlywere run.robot.jsonhash remained identical after policy-only update (6fdb2b039a5f6df83418d46de9129dd2bfa742c1e7623349614189697d2b2b5c). No training process or source training code was changed.
Numerical/browser checks
Real Chromium + MuJoCo WASM + ONNX Runtime WASM; not a mocked controller. The initial 3-second independent trials were reset separately:
| Trial | World X displacement | Yaw change | Fall |
|---|---|---|---|
| stand | −0.0021 m | 0.0002 rad | none |
| forward +0.4 | +0.6953 m | −0.0150 rad | none |
| backward −0.3 | −0.5985 m | approximately 0 rad | none |
| left +0.6 | −0.0361 m | +1.6520 rad | none |
| right −0.6 | −0.0688 m | −1.7249 rad | none |
Three deterministic 25-second sequences also pass: stand 5 s → forward 3 s → rest 2 s → backward 3 s → rest 2 s → left 3 s → rest 2 s → right 3 s → rest 2 s. No falls or in-sequence resets; root positions remain within |x|, |y| < 1.9 m (arena inner half-extent 2.25 m). Representative local forward/backward displacement: +0.7551 / −0.5143 m; turns +1.5705 / −1.5989 rad; minimum root height 0.7622 m; PD saturation count 0/120000 per sequence. Identical trials are repeatability evidence, not domain-randomized robustness evidence.
A four-second realtime trial advances at approximately 1.02× wall time (acceptance range 0.8–1.15), consistent with the 50 Hz scheduler. The bound is measured for simulation/control, not a GPU rendering FPS promise.
Keyboard W/S/A/D commands match manifest limits; key-up clears them. J/H are shown but disabled while no weights exist. Artificial side tilt triggers fall detection; Space clears fall, history and FSM. Microduck → Microban → RP1 transitions are exercised. Screenshot: Playwright test-results/rp1-third-robot-*/rp1.png (generated, ignored by git).
Interface checks
- Unit tests cover term-major/frame-major history, first-sample backfill, live debug reads without history pushes, reset, policy order, named gains, torque-speed limits, atomic invalid-action handling, independent skill defaults/history, load-failure session release, stop gating, timeout, blend, movement interruption and unavailable input suppression.
- The episode browser fixture explicitly reuses the walk ONNX to exercise real stop gating, two sessions/reset/return ownership and movement interruption; it is not a dance/wave implementation or a quality test of those future motions.
- Original BeyondMimic policy observations include motion-preview/anchor/reference terms absent from this runtime. Those weights require a real reference adapter or compatible export before enabling the reserved skills.
Final results
validate:packages: all four manifests (three families), resource paths and provenance hashes pass.npm test: 47/47 unit tests pass.format:checkand production build pass (the existing MuJoCo browser-externalmodulewarning remains non-fatal).- Full browser regression: 36/36 pass, including all nine unchanged Microduck deterministic baselines and Microban long walking trials.
- After restricting episode blending to primary-owned joints and enhancing tests, the final 6/6 affected browser tests were rerun successfully: three-family deployment, original Microban mounted-IMU, RP1 keyboard/fall/switching, RP1 3×25 s/realtime, real stop-gated episode/interruption, and 20 ownership transitions now cycling through all three robot families.
- In that final ownership run, every retired model/data is deleted. RP1 listener count remains 15, scene children 5, renderer geometry/texture counts plateau at 69/2; no monotonic leak is observed.
Reproduce
npm --prefix app run validate:packages
npm --prefix app test
npm --prefix app run build
npm --prefix app run format:check
npm --prefix app run test:e2e
Playwright JSON output contains the rp1-preview-trials.json, rp1-three-sequences.json and rp1-realtime.json attachments. A full-suite run caught stale single-frame diagnostic reads after introducing history; write({advance:false}) was corrected to sample live state without committing a history push, and the original Microban mounted-IMU regression then passed.
Limits
Intermediate weights only; no statement of training convergence, real-robot readiness, recovery-policy availability or arbitrary BeyondMimic compatibility. Random delays, domain randomization and training disturbances are omitted. Motion-transition blends enforce target continuity, not universal dynamic safety. RP1 description licensing remains unprovided in this checkout; local evaluation only, not clearance for public asset redistribution.