[PERF] Split rigid collision BVH into static + dynamic subsets (RPL multi-depth)

[Kashu7100]

Description

Follow-up to #2867. That PR added the static-rebuild-skip + shared-BVH-across-envs fast paths, but they only engage when every link in the solver is fixed (maybe_static = all(link.is_fixed for link in solver.links)). The common RL case — one moving robot on a large static terrain in a single rigid solver — never qualifies: the robot's movable links make the whole solver non-static, so the single combined collision BVH (terrain + robot) is rebuilt over every face each step. The static terrain, which is the bulk of the faces, is re-fit every step for nothing.

This PR decomposes the rigid solver's collision faces into two compacted BVHs by owning-link fixedness (RaycasterSensor._partition_collision_faces):

• static subset (faces on fixed links — terrain / walls): maybe_static, built once, then skipped + shared across envs (the dominant per-step cost for one robot on a big static terrain).
• dynamic subset (faces on movable links — the robot): rebuilt each step, but the rebuild + radix sort now scale with the robot's face count, not the whole scene.

The two are cast separately and merged via the existing is_merge path (closest hit wins), so the result is identical to one combined BVH. This is the "multi-depth" decomposition from RPL (arXiv:2602.03002): cast the dynamic robot and static terrain meshes separately and reuse the static acceleration structure across timesteps and environments.

Implementation

• Each BVH is built over a compacted face subset. A face_ids array maps a BVH leaf slot → global face index; bvh_ray_cast remaps after reading the morton-code primitive id, and update_aabbs iterates the subset. n_triangles in bvh_ray_cast now derives from the morton-codes shape (the BVH's own leaf count) instead of the solver-global face count.
• The existing maybe_static/needs_rebuild skip and the AABB-derived shared_across_envs test are already per-entry, so they apply to each subset unchanged: the static subset's GEOMETRY subscriber fires only on an explicit set_pos/set_quat (e.g. re-randomized terrain), never on physics-driven robot motion — so it stays skipped + shared while the robot subset rebuilds.
• A pure-static or pure-dynamic solver yields a single subset with identity face_ids, i.e. the previous single-BVH behavior — bit-identical.
• kernel_cast_ray (viewer pick) and the viewer plugin thread an identity face_ids over the full mesh.

Motivation and Context

Profiling perceptive whole-body imitation (torso depth camera, 64×36 rays, 18.5k-face terrain + G1-29DoF, RTX 3080): the combined-BVH rebuild dominated env.step and #2867's static-skip never engaged because the robot shares the solver. With the split, a moving-robot-on-terrain scene builds 1 static (skipped + shared) + 1 dynamic BVH; depth output is unchanged.

Per-step (measured on the pre-refactor branch): 1024 envs 1037 → 706 ms/step (1.47×); the win grows with env count × terrain face count (the eliminated static rebuild scales with both).

How Has This Been Tested?

Environment: single RTX 3080, CUDA backend.

• New test tests/test_sensors.py::test_raycaster_static_dynamic_bvh_split (n_envs ∈ {0, 2}): asserts the split structure (one static + one dynamic collision BVH; static shared_across_envs), that the merge reports the closer of static/dynamic as a movable box enters/leaves a ray, and that the static BVH stays skipped (needs_rebuild False) across a dynamic-only move.
• Existing raycaster/lidar suite unchanged: the single full-mesh path is the identity-face_ids case (bit-identical); mixed scenes now exercise the two-BVH merge.
• End-to-end: a real depth-camera perceptive task builds + steps with the expected static/dynamic split and unchanged depth.

Checklist

• I read the CONTRIBUTING document.
• I tagged the title correctly ([PERF]).
• I added a test (test_raycaster_static_dynamic_bvh_split).
• All existing tests pass (CI).

🤖 Generated with Claude Code

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 48736869ad

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[Milotrince]

Rebased onto current main and fixed a correctness bug!

[github-actions]

🔴 Benchmark Regression Detected ➡️ Report

[duburcqa]

@codex review

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 47675421c3

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[duburcqa]

Rebasing onto today's main turned into a near-rewrite: the per-component repack, the distances-only raycaster and the tactile single-tree BVH queries all landed in between, and on top of that I added compile-time remap gating, a combined-BVH contract for the tactile sensors and a pose-based rebuild-skip gate. Rather than ship all of that under your name, I moved the final version to #3078 (with credit back here) and reset this branch to your original commit.