Python: per-environment MJCF insertion + collision-capacity override - #16
Closed
haixuanTao wants to merge 2 commits into
Closed
Python: per-environment MJCF insertion + collision-capacity override#16haixuanTao wants to merge 2 commits into
haixuanTao wants to merge 2 commits into
Conversation
…rride insert_mjcf gains an env= kwarg (default 0, fully backward compatible): the robot, its auto-floor and the camera-fit AABB all target rbd_world_mut(env), and viewer registration (visual meshes, camera, light) only runs for env 0 since the viewer draws environment 0 — batch environments are physics-only. add_environment() was already exposed, so one robot per env is now reachable from Python. Why: loading N robots into env 0 is quadratic — every contact-constraint slot stores a dense M^-1 column sized to the WHOLE env's DOF count (contact_cons_col_cap = num_mbs * 192 * dofs_cap), which at 2048 robots requests ~58 GiB twice and caps single-env packing at 96 robots (121 MiB, just under wgpu's 128 MiB binding default). One robot per env keeps dofs_cap at the robot's own DOFs and scales linearly with batches. Also adds NexusState::set_rbd_collisions_capacity(): the 4096-pairs/env default is sized for one busy scene, not thousands of small envs — pair-keyed workspaces bind capacity x envs x sizeof(manifold) ~ 9.4 GiB at 2048 envs for a scene that generates ~7 pairs per env. Lowering it to 64 also halved wall-clock at 256 envs (17.6k -> 36.1k env-steps/s). Measured (RTX 5090, headless WebGPU, 12-DOF LeRobot biped, passive, 200-step window): 2,048 envs = 138,640 env-steps/s; linear from 256 (36k). Note batched finalize requires identical collider counts per env. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
haixuanTao
force-pushed
the
feat/per-env-mjcf-upstream
branch
from
July 16, 2026 12:17
e839c74 to
0447bf0
Compare
sebcrozet
requested changes
Jul 16, 2026
sebcrozet
left a comment
Member
There was a problem hiding this comment.
That looks good, thank you. Just a few comments that got strangely merged.
Comment on lines
+157
to
+163
| /// Sets the number of rigid-body solver steps advanced per | ||
| /// [`NexusPipeline::simulate`](crate::pipeline::NexusPipeline::simulate) call (default 1). Acts as a simulation-speed control. | ||
| /// Overrides the per-environment collision-pair capacity used when the | ||
| /// GPU rigid-body state is (re)allocated at `finalize`. The default (4096) | ||
| /// is sized for one busy scene, not thousands of small batched envs — | ||
| /// pair-keyed workspaces scale as `capacity x num_envs x sizeof(manifold)`, | ||
| /// which at 2048 envs binds ~9 GiB unless this is lowered. |
Member
There was a problem hiding this comment.
The comments got mixed up (the first two sentences should apply to set_rbd_steps_per_frame).
Contributor
Author
There was a problem hiding this comment.
Good catch — fixed in 1aa4958: the set_rbd_steps_per_frame doc is back on its own function.
Comment on lines
+266
to
+276
| /// Loads a MuJoCo MJCF scene into environment 0 as multibodies, registering | ||
| /// its render shapes (and a sized floor) with `viewer`. Returns scene info | ||
| /// (suggested camera + whether the scene is Z-up). Call `finalize` after. | ||
| #[pyo3(signature = (viewer, scene_path, render_colliders=false))] | ||
| /// Per-environment collision-pair capacity (default 4096). Lower this | ||
| /// before `finalize` when batching many small environments — pair-keyed | ||
| /// GPU workspaces scale with `capacity x num_envs`. | ||
| fn set_rbd_collisions_capacity(&mut self, capacity: u32) { | ||
| self.0.set_rbd_collisions_capacity(capacity); | ||
| } | ||
|
|
||
| #[pyo3(signature = (viewer, scene_path, render_colliders=false, env=0))] |
Member
There was a problem hiding this comment.
The comments got mixed up here too.
Contributor
Author
There was a problem hiding this comment.
Fixed in 1aa4958 — moved the insert_mjcf doc back onto insert_mjcf (and updated its wording for the new env parameter).
Review feedback: the set_rbd_steps_per_frame doc was merged onto set_rbd_collisions_capacity, and insert_mjcf's doc onto the Python set_rbd_collisions_capacity binding. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Member
|
This will be merged as part of #36 |
sebcrozet
added a commit
that referenced
this pull request
Aug 29, 2026
* fix(rbd): apply the per-batch stride to collider_parent reads in the narrow phase Replaces #21 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * feat: per-environment collision-pair capacity override Replaces #24 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * feat(rbd): make the narrow-phase contact prediction distance configurable Replaces #28 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * fix(rbd): thread the configurable prediction distance through the brute-force broad phase Completes #28 * feat(python): per-environment MJCF insertion Replaces #16 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * fix(python): drop the duplicated collisions-capacity setter and pass the RbdCoupling to insert_rigid_body_in Completes #16 * feat(python): per-step MJCF actuator control + multibody state readback Replaces #12 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@1ms.ai> * fix(python): gate multibody control/readback on dim3, add the missing PyArray2 import Replaces #12 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * fix(rbd): decode the SoA link workspace for the multibody readback and drop the stale set_gravity copy Completes #12 * perf(rbd): dedupe shared TriMesh uploads in from_rapier Replaces #19 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * perf(rbd): optional GPU contact reduction, merging per-pair manifolds to <=4 points Replaces #17 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * fix(rbd): pass the prediction distance to manifold_reduction in the contact-reduction kernel Completes #17 * perf(rbd): flat 1-D narrow-phase dispatch, packing warps across batches Replaces #21 Co-Authored-By: Haixuan Xavier Tao <tao.xavier@outlook.com> * fix(rbd): restore the contacts capacity binding and import atomic_load_u32 for the flat dispatch Completes #21 * fix(rbd): drop the stale 2mm PREDICTION constant reintroduced by the flat-dispatch port Completes #21 * fix mpm feature-gating * feat(rbd): expose dof_state_mut, links_static, joint_constraints and link_of_body * feat(rbd): env-reset primitives, GPU motor scatter, contact sensors, actuator delay, encoded step, substep-refresh cadence and per-DoF armature/frictionloss * fix(rbd): guard against implicit-coriolis drifting from the batch_indices uniform * feat(rbd): cluster contact manifolds by normal, matching rapier, with a tunable threshold * feat(rbd): model multibody joint frictionloss as a constraint instead of a force * feat(rbd): seed per-DoF joint friction from rapier's Multibody::frictions * refactor(rbd): read the contact prediction distance from RbdSimParams instead of a dedicated uniform * chore: cargo fmt * refactor(rbd): move the contact merge cosine into RbdSimParams * refactor: move read_multibody_links onto NexusState and drive every env from control_multibody_motors * test(rbd): add a headless many-small-environments step-timing harness * revert(rbd): drop the flat 1-D narrow-phase dispatch Measured 7-21% slower on Metal. * chore: cleanup comments * fix: gate control_multibody_motors on dim3 so the 2D build still compiles * fix instability in joint-ball3 demo * chore: remove debug test files * chore: cleanups * fix(rbd): build the bench harness without the metal feature and only in 3D * fix(rbd): silence the clippy needless-borrow and unnecessary-mut lints * fix(rbd): split the joint-constraint back-solve into its own dispatch to fit 8 storage buffers * test(rbd): keep the bench harness under wgpu's default buffer-size limit * fix(rbd): split the batched env reset into pose and DoF passes to fit 8 storage buffers * chore: switch to the published rapier version * fix: make all envs share the same RbdSimParams * chore: clippy fixes --------- Co-authored-by: Haixuan Xavier Tao <tao.xavier@outlook.com> Co-authored-by: Haixuan Xavier Tao <tao.xavier@1ms.ai>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
insert_mjcfgains anenv=kwarg (default 0 — fully backward compatible), making one-robot-per-environment reachable from Python: the robot, its auto-floor and the camera-fit AABB targetrbd_world_mut(env), while viewer registration (visual meshes, camera, light) stays env-0-only since the viewer draws environment 0 — batch environments are physics-only.add_environment()was already exposed.Why
Packing N robots into env 0 is quadratic: every contact-constraint slot stores a dense M⁻¹ column sized to the whole environment's DOF count (
contact_cons_col_cap = num_mbs × 192 × dofs_cap). At 2,048 robots that requests ~58 GiB twice (columns + Jacobians), and single-env packing caps at 96 robots — 121.5 MiB, just under wgpu's 128 MiB binding default. One robot per env keepsdofs_capat the robot's own DOFs and scales linearly withbatches, which is the layout the batched solver is designed around.Second scaling fix: new
NexusState::set_rbd_collisions_capacity(). The 4,096-pairs/env default is sized for one busy scene, not thousands of small envs — pair-keyed workspaces bindcapacity × envs × sizeof(manifold)≈ 9.4 GiB at 2,048 envs, for a scene generating ~7 pairs/env. Lowering it to 64 also halved wall-clock at 256 envs (17.6k → 36.1k env-steps/s): the sweep kernels were paying for the slack.Measured (RTX 5090, headless WebGPU, 12-DOF LeRobot biped, passive, 200-step window, one sync)
Linear scaling; cross-checks against the zealot training stack's physics-only rate at 2,048 envs (~127k on the same engine family).
Notes
finalizerequires identical collider counts in every environment — worth a friendlier error eventually.🤖 Generated with Claude Code