roadmap: g1_motion_tracking multi-backend enablement — newton CUDA graph, genesis cold-start, isaac worker fixes

[TATP-233]

Background

FlashSAC g1_motion_tracking now runs on five backends (mujoco, motrix, mjwarp, newton, genesis) after #1603, with training metrics at or above the fastsac baseline. Three known gaps remain, and closing them spans three repositories: Motphys/UniLab (this repo, owner configs/registration), unilabsim/unisim (backend adapters), unilabsim/unilab_rl (algo runners/collectors). This issue tracks the cross-repo roadmap.

Work items

1. Newton CUDA graph acceleration (unisim → unilab)

The newton backend reaches ~35k sps vs mjwarp's ~94k on the same task. Unlike mjwarp — where CUDA graphs are automatic and transparent (unisim/backend/mjwarp/backend.py, _cuda_graph_eligibility(): CUDA + driver ≥ 12.4 + mempool, eager fallback) — NewtonBackend has no graph support at all, and newton.SolverMuJoCo (1.5.1) exposes no graph parameter.

• unisim: implement CUDA graph capture in NewtonBackend, following the mjwarp eligibility/fallback pattern. Constraints identified:
  • _physics_substep swaps state/state_out pointers each step → capture the full nsteps loop (or parity double-graphs)
  • set_state (reset) allocates fresh warp arrays per call → keep reset outside the graph unless fixed-capacity scratch is added (mjwarp _RESET_SCRATCH_CAPACITY pattern)
  • capture must happen after materialize() capacity calibration rebuilds state
  • per-step capacity-check host sync (backend.py:623-632) limits graph benefit; consider periodic checking (behavior change, needs review)
• unilab: optional owner knob (newton_use_cuda_graph) in src/unilab/base/base.py + backend_factory.py forwarding; document in newton owner YAMLs
• Validate with flashsac g1_motion_tracking full-budget training (target: sps uplift, reward parity with the 26.50 / best 34.15 eager result)

2. Genesis cold-start tick-0 timeout — root fix (unilab_rl → unilab)

First genesis run always hits TimeoutError: collector inference tick 0 (30s): collector env construction — including genesis kernel JIT (all of it inside scene.build, no lazy compilation left) and the first reset — happens inside the tick-0 timing window (uni_rl/offpolicy/double_buffer_runner.py:555, worker.py:232→325). The disk cache (~/.cache/quadrants/qdcache/) fixes second runs, but even warm builds at 2048 envs can exceed 30s (per-process kernel module loading). The existing learner-side 1-env probe env does not warm the collector's cache: genesis forks kernel sets by n_envs (PARA_LEVEL.PARTIAL for n_envs≤1 vs ALL otherwise, genesis/engine/scene.py:967).

Current state: timeout override (training.inference_request_timeout_sec) is only a mitigation, and flash_sac doesn't even forward it yet (unilabsim/unilab_rl#26; fast_td3 is missing it too).

• unilab_rl (root fix): collector-ready handshake — collector signals after init_state; learner starts tick timing only after ready, so 30s bounds steady-state ticks, not backend cold start. Merge with unilab_rl#26 (timeout forwarding for flash_sac/fast_td3)
• unilab (mitigation, doable now): set GS_PARA_LEVEL=2 for the learner probe env when sim_backend=genesis so the probe compiles the same kernel set as the collector, moving cold JIT into the untimed runner-construction phase (src/unilab/scripts/train_offpolicy.py)
• After the unilab_rl handshake lands: re-validate genesis at the default 2048 envs (currently validated at 512 envs, 12.8M steps: final 23.94 vs baseline 12.33 @10m steps)

3. isaacgym/isaacsim worker state fixes (unisim → unilab)

• unisim: fix worker state publish/reset path (isaacgym/isaacsim: broken body-state publish / reset for motion-tracking workloads unilabsim/unisim#141 — episode length stuck at 1.0-1.6; evidence attached there). Not fixed in unisim-core 1.5.1
• unilab: after the unisim fix, probe both backends on g1_motion_tracking, register them in src/unilab/tasks/motion_tracking/g1/__init__.py, add owner YAMLs, and run full-budget flashsac training validation
• Regenerate the support matrix and update tests/envs/test_env_configs.py backend list

References

• newton/genesis backends: implement SimBackend.get_motion_body_ids (blocks motion-tracking tasks) unilabsim/unisim#137 — newton/genesis get_motion_body_ids (fixed in unisim-core 1.5.1 via feat: implement get_motion_body_ids on newton/genesis, fix subprocess offset (#137) unilabsim/unisim#140)
• isaacgym/isaacsim: broken body-state publish / reset for motion-tracking workloads unilabsim/unisim#141 — isaacgym/isaacsim worker state publish/reset broken
• flash_sac double-buffer builder does not forward inference_request_timeout_sec unilabsim/unilab_rl#26 — flash_sac does not forward inference_request_timeout_sec
• feat: add flashsac g1_motion_tracking task owners and newton/genesis backends #1603 — flashsac g1_motion_tracking owners + newton/genesis registration + unisim-core 1.5.1 bump

[TATP-233]

Newton CUDA graph full-budget validation is complete on RTX 4090.

• unisim implementation: feat: add Newton CUDA graph execution unilabsim/unisim#144, head 3402c8f
• UniLab owner knob: feat: add Newton CUDA graph owner knob #1605, head 800a8f7c
• Graph run: FlashSAC g1_motion_tracking, 2048 envs, 25,000 iterations, seed 1, AMP, env.newton_use_cuda_graph=true, no playback
• Eager baseline: 34,953.0 final SPS, 1,553.85 s training wall time, 26.4967 final / 34.1479 best mean reward
• Graph run: 51,928.9 final SPS, 998.03 s training wall time, 29.1227 final / 34.2186 best mean reward
• Result: +48.57% final SPS, -35.77% training wall time; reward parity holds
• Run status: completed, 25,000/25,000 iterations, 51,400,704 env steps; CUDA mempool enabled and no graph fallback warning

Evidence comments: #1605 (comment) and unilabsim/unisim#144 (comment)

Separately, the Genesis learner-probe mitigation is now review-ready in #1606 with local make test-all, docs checks, Sphinx build, and full current-head remote CI all passing.