Mesh non-collision loss optimization

[qianl-nv]

Summary

Batch mesh collision SDF queries to reduce solver time from 55s to 4s for the droid kitchen scene.

Detailed description

• The baseline runs sphere to SDF serial per-candidate in compute_no_overlap_loss_mesh resulting in one warp launch per candidate. While the SDF kernel itself runs fast (0.75ms), the serial multi launch results in NoOverlapMESH taking 55ms per iteration (97% of each interation time).
• Replace per-candidate mesh queries with one batched Warp launch. compute_no_overlap_loss_mesh rewritten as a fully tensorized (B, P) / (B, S) pipeline (batch × pairs × spheres).
• Remove unnecessary host sync points — .item() when building yaw tensors from Python lists; if not active_pair.any(): continue forcing GPU→CPU sync every iteration.

Tests

• B=4, mixed overlap/yaw; loss and ∇positions match baseline serial implementation
• For existing droid kitchen scene, validation pass rate for serial vs batched implementation is identical.
• Manual run droid_pick_and_place_lightwheel_kitchen.yaml and layout looks correct over 20 resets.

Command to run:
/isaac-sim/python.sh isaaclab_arena/evaluation/policy_runner.py --viz kit --policy_type zero_action --num_episodes 20 --disable_fabric --device cpu --env_graph_spec_yaml isaaclab_arena_environments/kitchen_bench/droid_pick_and_place_lightwheel_kitchen.yaml

with solver logging enabled should see
[RelationSolver] solve: 3715.2 ms | batch=50 | objects=3 optimizable + 3 anchors | no-overlap pairs=12 | iters=600 (6.19 ms/iter)
vs baseline
[RelationSolver] solve: 56006.8 ms | batch=50 | objects=3 optimizable + 3 anchors | no-overlap pairs=12 | iters=600 (93.34 ms/iter)

[arena-review-bot]

🤖 Isaac Lab-Arena Review Bot

Summary

This PR rewrites compute_no_overlap_loss_mesh from a per-env / per-candidate serial loop into a fully tensorized (B, P) / (B, S) pipeline with a single Warp SDF launch, and removes host-sync points — a solid, well-motivated optimization (55s → 4s). The dense-query-then-mask trade-off is documented, and the added regression test checks both loss and position-gradient equivalence against a serial oracle across a mixed overlap/yaw batch, which is exactly the right coverage. My only concern is a pair of profiling toggles that look like benchmarking leftovers.

Findings

🟡 Warning: relation_solver_params.py:58 — profile default flipped False → True, so every solve() now prints a timing summary for all users. Unrelated to the batching work; likely a benchmarking leftover to revert.

🟡 Warning: arena_env_builder.py:101 — the default builder path (placer_params is None) now hard-codes profile=True, enabling timing prints for every environment built through the builder. Same question — intended or leftover?

Test Coverage

Good. test_batched_mesh_loss_matches_test_only_serial_oracle compares the batched implementation against the retired serial oracle (kept in the test file) for both loss values and optimizable_positions gradients, over a B=4 batch mixing overlap and yaw. Consistent with the file's existing warp/torch test style — no SimulationApp, so the inner/outer sim pattern isn't required.

Verdict

Minor fixes needed — confirm/revert the two profile=True toggles; the core optimization looks correct and well-tested.

[greptile-apps]

Greptile Summary

The PR batches mesh collision SDF work across candidates to replace many serial Warp launches with one dense query and adds a serial-reference equivalence test.

• Tensorizes pair positions, yaw transforms, broad-phase masks, sphere queries, and pair aggregation.
• Adds device-resident fixed-pose, fixed-yaw, and bbox-yaw metadata to the mesh-pair cache.
• Verifies batched losses and position gradients against the former serial implementation.
• Also enables solver profiling globally and in the environment builder's fallback configuration.

Confidence Score: 4/5

The PR appears safe to merge, with the non-blocking caveat that solver profiling should remain opt-in to avoid unsolicited timing output in normal runs.

The batched mesh-collision path preserves the investigated loss, gradient, cache, and yaw behavior, but the configuration changes enable profiling diagnostics for routine relation solves.

Files Needing Attention: isaaclab_arena/relations/relation_solver_params.py, isaaclab_arena/environments/arena_env_builder.py

Important Files Changed

Filename	Overview
isaaclab_arena/relations/no_overlap_mesh.py	Replaces per-candidate mesh queries with a batched tensor pipeline while preserving the checked position, yaw, broad-phase, loss, and gradient semantics.
isaaclab_arena/relations/mesh_pair_cache.py	Adds validated device-tensor cache fields for fixed poses, fixed yaws, and bbox yaw metadata.
isaaclab_arena/tests/test_mesh_collision.py	Adds a serial reference implementation and checks batched loss and gradient equivalence across mixed overlap, yaw, and fixed-obstacle cases.
isaaclab_arena/relations/relation_solver_params.py	Enables profiling by default, causing ordinary solver runs to emit timing diagnostics.
isaaclab_arena/environments/arena_env_builder.py	Explicitly enables profiling for fallback relation-solver parameters used during environment construction.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  A[Batch candidate positions and pair metadata] --> B[Compute batched yaw-aware AABB overlap]
  B --> C[Transform all sphere centers into obstacle frames]
  C --> D[Single dense multi-mesh SDF launch]
  D --> E[Mask inactive pairs]
  E --> F[Scatter sphere penetration by pair]
  F --> G[Per-candidate no-overlap loss]

Loading

_{Reviews (1): Last reviewed commit: "Cleanup unused MeshPairCache data" | Re-trigger Greptile}

[qianl-nv]

self review 1