Skip to content

STATUS.md

Latest commit

 

History

History
51 lines (41 loc) · 8.11 KB

STATUS.md

File metadata and controls

51 lines (41 loc) · 8.11 KB
Raw

Isaac Lab integration — verification status

Native G1 demo — September 6, 2026

The researcher walkthrough was executed in Isaac Sim 5.1.0 on an L40S using the original Unitree locomotion checkpoint. 3 of 4 selected demo cases passed, with no simulator errors in the final suite: open doorway, automatic slider, and passive saloon-door passage succeeded; the closed latched-door attempt failed and the robot fell. This is a canonical-USD locomotion integration demo, not PPO training, a human reference, or an all-door benchmark. Native receipts identify the exact inputs and runtime.

The historical validation records below have their own scope and source revisions.

Written on an Apple-silicon Mac without an NVIDIA GPU: Isaac Sim / Isaac Lab cannot run here. This file lists exactly what was verified locally and what awaits the GPU run (task board I4).

Verified on this machine

what result
USD export rewrite (doorbench/export/usd.py): default prim = door root, Env (static) + Articulation (fixed base link, single tree), PhysX applied schemas (PhysxArticulationAPI, PhysxRigidBodyAPI, PhysxJointAPI, PhysxJointAxisAPI:angular/linear friction efforts, PhysxCollisionAPI, PhysxConvexHullCollisionAPI, PhysxMimicJointAPI with rel referenceJoint), /PhysicsScene outside the default prim, per-joint doorbench:* metadata regenerated 1000 doors in 37 s (scripts/generate_dataset.py --formats mjcf,urdf,usd,json); MJCF / URDF / thumbnails byte-identical; n_signed_off = 1000 (QA now also requires usd_rl_opens)
Physics-mapping fixes after parity round 1 (2026-09-05, docs/ISAAC_LAB.md "MuJoCo → PhysX parameter mapping"): per-axis friction API moved from the ignored rotX/transX instance to angular/linear (readback was 0.0 on all 1000 doors), legacy physxJoint:jointFriction authored 0 (was load-dependent, 7-10× on the turnstile columns), DOOR_RIGID_PROPS.max_angular_velocity 100 deg/s → 5729.58 deg/s (= 100 rad/s; every leaf had plateaued at 1.9 rad/s) and authored on every link, MJCF position servos of spring-less joints folded into the PhysX drive (doorbench:servo_in_drive), rising-hinge gravity torque in doorbench:rl["rise_coupling"] for the locked riser, armature on both APIs; runner / DoorState read the friction back and write it through Isaac Lab when PhysX disagrees regenerated 1000 doors (MJCF byte-identical); static validator 1000/1000 with the new checks (instance names, legacy coefficient 0, maxAngularVelocity ≥ 1000 deg/s, servo drive consistency, rise coupling); tests/test_usd_physics_mapping.py (51 tests, exports one door per family without assets/)
New canonical door_rl.usda (8 links / 7 joints for every door, doorbench:rl metadata) 1000/1000 written; slot histogram in assets/usd_validation.json
scripts/isaaclab/validate_usd_static.py over all 1000 doors, both files 1000/1000 pass — full: 3 618 joints, 4 650 rigid bodies, 21 719 colliders, 5 347 mesh references resolved, 0 warnings; rl: 7 000 joints, 8 000 rigid bodies, 20 600 colliders, 57 warnings (doors that cannot open by spec: jammed / interlocked / child-locked, and 35 doors without a handle site → leaf point used). Checks: stage metadata, single articulation root, fixed base, tree connectivity, mass/inertia > 0, joint frames consistent through body0/body1 (anchor & axis), limits, drives (gains, spring targets vs model.json within 1e-3), friction efforts == IR Coulomb values, collision APIs + physics materials, mesh references, JSON attributes, RL slot consistency
pytest -q tests/test_doorbench.py pass (6 tests)
pytest -q tests/test_mujoco_import.py on the regenerated dataset pass (261 tests: MJCF/URDF still byte-identical, QA sign-off intact)
pytest -q tests/test_isaaclab_ext.py (new) pass: static validation of one door per family + 20 random, RL structure / meta, doors index & easy-100 curation, hand USD, py_compile of the extension + scripts, offline API-name checklist
python -m py_compile of every new file pass
scripts/isaaclab/check_api_names.py 130 Isaac Lab / rsl-rl symbol references in 22 files, all present in the Isaac Lab v2.3.2 reference list; --source <IsaacLab v2.3.2 clone> --source <rsl_rl v3.1.2 clone> additionally resolves every symbol in the real tree and checks the keyword arguments of 128 config-class / function calls against the fields defined there (2026-09-05, all pass)
Isaac Lab v2.3.2 API audit (after the train stage crashed on dump_pickle) isaaclab.utils.io lost dump_pickle/load_pickle (isaaclab 0.47.0, security) -> local replacement in scripts/isaaclab/_common.py; RslRlVecEnvWrapper.get_observations() returns a TensorDict (no (obs, extras) tuple); the policy normalizer is policy.actor_obs_normalizer (play.py export fixed); empirical_normalization is deprecated in rsl-rl 3.1 (dropped from the PPO cfgs); the runner is picked by agent_cfg.class_name; env_cfg.log_dir is set. Everything else we use (ArticulationCfg.articulation_root_prim_path, MultiUsdFileCfg, actuator / schema / sensor cfg fields, mdp.*, set_camera_view, scene.stage / env_prim_paths, sim._app_control_on_stop_handle) exists in v2.3.2
isaaclab/cloud/run_all.sh validation-first: the validate stage runs the Isaac parity runner over ALL doors when the repo has one (auto-detected scripts/isaaclab/*parity*.py, override PARITY_RUNNER), else validate_usd_isaacsim.py --all; train / hero / eval are OFF unless TRAIN=1 (skipped stages write STAGE_<name>_EXIT=skipped). Stage output is teed to logs/stage_<name>.log and a Python traceback counts as a failure: the 2026-09-05 pod run printed STAGE_train_EXIT=0 right after the dump_pickle ImportError because the simulator shutdown masked the exit status; _env.sh's set -e is switched off so a failed stage no longer aborts the pipeline. Dry-run with a fake launcher (4 scenarios) on 2026-09-05
isaaclab/doorbench_isaaclab/data/gantry_hand.usda generated + validated (6 DoF, fixed base, drives)

Not executed (needs the GPU box; run bash isaaclab/cloud/validate.sh first)

  • isaaclab/cloud/setup.sh now wraps scripts/pod_bootstrap.sh, which was executed end to end on a RunPod L40S on 2026-09-04/05 (Isaac Sim 5.1.0 wheels via uv, Isaac Lab v2.3.2, ISAACLAB_IMPORT_OK); see docs/RUNPOD.md. Dockerfile: still untested.
  • scripts/isaaclab/validate_usd_isaacsim.py: Isaac Sim import of all doors (spawn, settle, actuate). This is the first thing to run; it tells whether PhysX parses the articulations as intended (fixed base, joint frames, friction efforts, mimic joints).
  • doorbench_isaaclab environments (DoorBench-Open-Hand-v0, DoorBench-Open-G1-v0): never instantiated. Risk points, in order of likelihood: (1) articulation_root_prim_path="/Articulation" with MultiUsdFileCfg (fallback: remove the argument), (2) the 0-dim DoorMechanismAction (fallback: interval event), (3) contact-sensor filter regexes for the G1, (4) G1 body-name candidates for the hands (.*_palm_link), (5) rsl-rl config fields if the installed Isaac Lab is not v2.3.x.
  • Training quality: reward weights follow the benchmark events + Isaac Lab's G1 regularisers but were never tuned.
  • record_hero.py (viewport render via env.render() with --enable_cameras), eval_all_doors.py.
  • PhysX semantics that only a run can confirm: mimic-joint gearing units. Joint friction is no longer an unknown: Isaac Sim 5.1 reads the PhysxJointAxisAPI efforts only from the angular / linear instance (round 1 authored rotX / transX and read back 0.0 everywhere); isaac_parity.py and DoorState now compare the read-back with the IR and write the efforts through write_joint_friction_coefficient_to_sim if PhysX disagrees, so a parser regression shows up as a structure error instead of a frictionless door.

How to report back

After bash isaaclab/cloud/validate.sh and one train.sh run, paste assets/usd_validation_isaacsim.json's summary and the first 20 lines of the training log into the task board (I4); fixes will be small and local.