Work: Investigate low MotrixSim physics throughput at 8k batched rollout #119

Work type

benchmark

Area

motrix

Problem

Benchmark evidence shows that motrixsim batched physics throughput is materially below mujoco.rollout on the same workload at num_envs=8192 and steps=20.

This gap is large enough to affect backend selection and training throughput expectations for the Motrix path.

Reproduction commands:

uv run python benchmark/benchmark_mujoco_vs_motrix.py --xml src/unilab/assets/robots/g1/scene_flat.xml
uv run python benchmark/benchmark_mujoco_vs_motrix.py --xml src/unilab/assets/robots/go2/scene_flat.xml
uv run python benchmark/benchmark_mujoco_vs_motrix.py --xml src/unilab/assets/robots/go1/scene_flat.xml

Observed output snippets:

mujoco_rollout_multi | 8192 | 188176.2 | 0.8707
motrixsim            | 8192 |  24637.7 | 6.6500

mujoco_rollout_multi | 8192 | 442776.2 | 0.3700
motrixsim            | 8192 |  99133.8 | 1.6527

mujoco_rollout_multi | 8192 | 435931.1 | 0.3758
motrixsim            | 8192 | 147402.1 | 1.1115

Deliverable

• Root-cause analysis for where time is going in the motrixsim benchmark path.
• Backend or adapter-layer optimization, or an explicit documented limitation if the gap is fundamental.
• Before/after benchmark artifact for g1, go2, and go1 under identical settings.

Definition of done

• Reproduce the baseline gap with benchmark/benchmark_mujoco_vs_motrix.py.
• Identify the main hotspot(s) with profiler or instrumentation evidence.
• Land fixes in the backend or adapter layer rather than script-only workarounds.
• Re-run and attach updated results for g1, go2, and go1 at num_envs=8192, steps=20.
• Document whether the remaining gap is expected, a bug, or a current backend limitation.

Dependencies and blockers

• None known yet. If the hotspot is inside MotrixSim internals, additional profiler support may be needed.

Proposed owner

Unassigned for now. Motrix backend owner to confirm.

Validation plan

• uv run python benchmark/benchmark_mujoco_vs_motrix.py --xml src/unilab/assets/robots/g1/scene_flat.xml
• uv run python benchmark/benchmark_mujoco_vs_motrix.py --xml src/unilab/assets/robots/go2/scene_flat.xml
• uv run python benchmark/benchmark_mujoco_vs_motrix.py --xml src/unilab/assets/robots/go1/scene_flat.xml
• If needed, add profiling or per-step timing around mtx.load_model, mtx.SceneData(...), warmup, and model.step(data) to isolate the bottleneck.