[hujc7]

Review Map

• A stacked PR's page shows the cumulative diff of its dependency chain; the "Exact changes" link pins that PR's own commit range.
• SHA links can go stale after a branch update; the copy on [Warp] Part 1/6: Reset only the masked environments and fix manager-based Warp environment construction #6646 is refreshed first.

PR	Status	Depends on	Exact changes
#6646 Part 1/6: mask-correct partial resets (draft)		—	—
#8783 Part 2/6: shared Warp RNG state (draft)		#6646	link
📌 #6607 Part 3/6: per-term capture, stable stepping parity (draft)		#8783	link
#8228 Part 4/6: Warp commands, Rough tasks, live term changes (draft)		#6607	link
#6582 Part 5/6: Shadow Warp twin, Allegro twin fixes (draft)		#6607	link
#7012 Part 6/6: --video on the Warp frontend (draft)		—	—

Summary

Warp environments step like the stable ones (physics-owned decimation, extras["final_obs"], per-step extras["log"] copies) and record their stages from the first call without an eager warm-up, bitwise equal to eager.
Capture is decided per term, so one term that cannot be recorded no longer forces its whole manager stage eager; Go2 Flat steps in 6.96 ms instead of 7.53 ms with Part 1 (8gpu L40, 4096 envs).

Description

Update (2026-10-22): now stacked on #8783 (the shared Warp RNG state, split out of #6646) and merged main 8c9b995cb95; this PR's own diff is unchanged.

Update (2026-10-20): merged #6646 and main 2dbb6146d75 (the default branch, renamed from develop; "main" below means main without this PR, measured at aa9bafef8a0). The Warp environments reach the physics backend through sim.backend_manager, since #8711 deprecated sim.physics_manager. After the merge, the experimental utils, managers and MDP parity tests (85), the frontend, scene, capture, decimation, final-obs, log and reset tests pass; Cartpole, Cartpole-Direct, Ant-Direct, Allegro-Direct and Go2 Flat Warp runs advance 16/16 RNG states, and every Cartpole and Go2 stage records one graph.

Update (2026-10-02): merged #6646 and develop 772358ac3b9, plus one review-round commit (ee9e8adbb9d): CapturedStage replaces WarpGraphCache, direct environments decide resets on the GPU, process_action takes torch actions, and a class term's reset is judged by its own annotation.

1. Changes

1. Capture: CapturedStage, @captured and eager() replace WarpGraphCache. A stage records on its first call without a warm-up, as main's physics capture does (Capture Newton CUDA graphs without advancing simulation #8058), replays afterwards, and records again after the simulation is reset or stopped (without this: CUDA error 700 after a rebind). Recordings pause the garbage collector and collect only the youngest generation afterwards. ISAACLAB_SYNC_DEBUG=1 raises on a hidden host sync.
2. Per-term capture: a @WarpCapturable(False) term runs eagerly at its place in the recording; @WarpCapturable(lambda params: ...) decides per configured instance. A class annotation covers the term's call; its reset is recorded unless the reset method is annotated.
3. Stepping like stable:
   • The environments call the managers in the stable order; ActionManager.process_action takes a torch action.
   • Manager-based environments reset by mask and derive indices once per reset, only for command, curriculum or recorder terms.
   • DirectRLEnvWarp decides on the GPU whether any environment must reset (a conditional graph node, CUDA 12.4+), so a step no longer synchronizes with the host.
4. Stateful event terms as classes (push_by_setting_velocity, apply_external_force_torque, reset_root_state_uniform, randomize_rigid_body_com): buffers allocated at creation, ranges read on every call. Fixes main's COM offsets accumulating per call and zero-range wrench events clearing permanent wrenches. An interval range replaced through EventManager.set_term_cfg now applies; main kept the old range.
5. Stable parity (fixes on main):
   • Physics-owned decimation: main steps physics and updates the scene 4 times per Go2 environment step, so contact history, air times, undesired_contacts and base_contact differ from stable.
   • extras["final_obs"] and Same-Step autoreset metadata: main ignores compute_final_obs, which the torchrl entry point sets.
   • extras["log"] copies: main returns views, so rsl_rl logs the last reset's value for every step of an iteration.
6. Breaking (experimental only):
   • Removed ManagerCallSwitch, ManagerCallMode, MANAGER_CALL_CONFIG, WarpGraphCache, ManagerBasedRLEnvWarp.step_warp_termination_compute and DirectRLEnvWarp.step_warp_action; direct tasks decorate their methods with @captured to keep them recorded.
   • ActionManager.process_action takes a torch tensor; SceneEntityCfg.joint_mask is renamed joint_mask_wp.
   • The Warp environments are still built from the stable task configurations (WarpFrontend.build_env swaps in the Warp MDP twins), so removing the switch drops no stable-configuration support.

2. Evidence

• Validation at 5ad9281530f (desk RTX PRO 6000): the experimental utils, managers, MDP, frontend, configuration-conversion, scene, capture, decimation, final_obs, log and reset suites pass. The capture tests compare eager and captured runs bitwise, including a simulation rebind mid-run. Cartpole, Cartpole-Direct, Ant-Direct, Allegro-Direct and Go2 Flat each step 300 to 1000 times; on Cartpole and Go2 Flat every manager stage records one graph.

• Step time (8gpu L40, 4096 envs, mean of 3 runs per arm, rotated across GPUs; measured on this PR before the last review fixes, which only change the reset path):

  Task	[Warp] Part 1/6: Reset only the masked environments and fix manager-based Warp environment construction #6646 (ms)	This PR (ms)
  Isaac-Cartpole	1.058	1.039
  Isaac-Cartpole-Direct	0.528	0.539
  Isaac-Ant	1.938	1.922
  Isaac-Velocity-Flat-UnitreeGo2	7.526	6.956

• Direct reset decision: resetting unconditionally every step is not slower (Cartpole-Direct 0.531 vs 0.538 ms), but it changes logged metrics on steps where nothing reset (Ant's Metrics/success_rate would read 0), so resets stay conditional as in stable.

• Decimation, Go2 Flat, 200 steps against the stable environment (earlier tip; the decimation tests pass at this tip): contact history, air times, undesired_contacts, base_contact and physics state are bitwise equal, feet_air_time within 3e-8 (reduction order).

3. Out of scope / follow-ups

• Direct Warp tasks cache .warp handles that go stale after an explicit hard sim.reset(); training never calls it.
• Stable-manager features that no supported Warp task uses still differ on the Warp managers: observation delay, runtime scale/clip edits, nested term-config resolution, recorder obs_buf, the GUI reset button and render_enabled.
• Faster manager data layouts (tiled reward, termination and command resets; an unrolled reward step; one observation post-processing launch): measured on desk, planned as a separate PR near the end of the series.
• OvPhysX on the Warp frontend ([WARP] Cleanup Part 4: Open the warp frontend to the OVPhysX backend #6580) is parked: lazily refreshed OvPhysX data is not refreshed inside recorded stages.
• Merging the stable and Warp managers into one implementation: proposal only, not part of this chain.

4. Merge note

#7012 changes the same lines of direct_rl_env_warp.py and manager_based_env_warp.py (video recorder hooks). Whichever of the two merges second keeps both additions.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• Breaking change (existing functionality will not work without user modification)

Release backport

• Backport this pull request to the active release branch after it merges into main

Screenshots

Not applicable.

Checklist

Docker and GPU tests run on demand. Push the commits you want tested, then
comment run-ci on the pull request.

• I have read and understood the contribution guidelines
• I have run the pre-commit checks with uv run isaaclab --format
• I have made corresponding changes to the documentation (API reference here; the capture guide is in Part 3)
• My changes generate no new warnings
• I have validated the changed behavior with focused coverage and applied the test-audit criteria to any test changes
• I have added a changelog fragment under source/<pkg>/changelog.d/ for every touched package (do not edit CHANGELOG.rst or bump package versions — CI handles that)
• I have added my name to the CONTRIBUTORS.md or my name already exists there

[hujc7]

run-ci

[hujc7]

run-ci

[hujc7]

run-ci

[hujc7]

run-ci