add settings to make training results deterministic by redfler · Pull Request #5449 · isaac-sim/IsaacLab

redfler · 2026-04-30T03:20:23Z

Description

This PR adds a deterministic training path and documentation for Isaac Lab RL workflows.

Added apps/isaaclab.python.headless.determinism.kit as a deterministic headless rendering experience.
Updated scripts/reinforcement_learning/rl_games/train.py to add opt-in --deterministic and use configure_seed(env_cfg.seed, args_cli.deterministic).
Updated docs/source/features/reproducibility.rst to document --experience isaaclab.python.headless.determinism.kit
and clarify that strict PyTorch determinism is currently exposed only for RL-Games.

Test command example:

./isaaclab.sh -p scripts/reinforcement_learning/rl_games/train.py --task Isaac-Cartpole-RGB-v0 --enable_cameras --headless --seed 42 --max_iteration 20 --deterministic --experience isaaclab.python.headless.determinism.kit

Fixes # (issue)

#3505 Non-reproducible training results in vision-based tasks with identical seeds

Type of change

Bug fix (non-breaking change which fixes an issue)

Screenshots

Before	After

Checklist

I have read and understood the contribution guidelines
I have run the pre-commit checks with ./isaaclab.sh --format
I have made corresponding changes to the documentation
My changes generate no new warnings
I have added tests that prove my fix is effective or that my feature works
I have updated the changelog and the corresponding version in the extension's config/extension.toml file
I have added my name to the CONTRIBUTORS.md or my name already exists there

isaaclab-review-bot

🤖 Isaac Lab Review Bot

Summary

This PR adds determinism support for vision-based RL training by: (1) calling configure_seed() in the rl_games training script to enable PyTorch deterministic algorithms, and (2) providing a new .kit experience file with RTX renderer settings that force synchronous rendering. The changes are minimal and focused, addressing a known reproducibility issue (GitHub #3505) for camera-based tasks.

Architecture Impact

train.py: The configure_seed() call affects only this specific training script. Other RL library wrappers (RSL-RL, SB3, SKRL) are NOT updated — users of those frameworks will still have non-deterministic vision training.
.kit file: New experience file extends isaaclab.python.headless with RTX determinism settings. No breaking changes to existing .kit files.

Implementation Verdict

Minor fixes needed — The core approach is sound, but there are gaps in test coverage, documentation, and consistency across training scripts.

Test Coverage

❌ No tests added. The PR description acknowledges this (checklist unchecked). For a bug fix claiming to resolve reproducibility issues, a regression test demonstrating determinism would be valuable — e.g., running 2 training iterations with the same seed and asserting identical loss/reward values. Without tests, the claim in the screenshots cannot be automatically verified.

CI Status

No CI checks available yet — cannot verify the changes don't break existing functionality.

Findings

🟡 Warning: scripts/reinforcement_learning/rl_games/train.py:207 — configure_seed hardcodes True for deterministic mode

configure_seed(env_cfg.seed, True)

The second argument enables torch.use_deterministic_algorithms(True), which can cause runtime errors for operations without deterministic implementations (e.g., certain scatter/gather ops on CUDA). This should either:

Be gated behind a CLI flag (e.g., --deterministic) so users can opt-in
At minimum, document which operations may fail

Looking at isaaclab.utils.seed.configure_seed, if the user hasn't explicitly requested determinism, forcing it could break existing workflows that previously worked.

🟡 Warning: Inconsistency across training scripts — only rl_games is updated
The RSL-RL (scripts/reinforcement_learning/rsl_rl/train.py), Stable-Baselines3, and SKRL training scripts are not updated. Users expecting determinism across all frameworks will be surprised. Either:

Update all training scripts consistently
Document that determinism is only supported for rl_games

🔵 Improvement: apps/isaaclab.python.headless.determinism.kit:168 — Missing newline at end of file

rtx.rtpt.lightcache.cached.enabled = false

File doesn't end with a newline. While minor, this violates POSIX standards and can cause issues with some tools.

🔵 Improvement: apps/isaaclab.python.headless.determinism.kit:12 — Package title doesn't reflect determinism purpose

title = "Isaac Lab Python Headless Camera"

The title says "Headless Camera" but this file is specifically for deterministic rendering. Should be "Isaac Lab Python Headless Deterministic Rendering" or similar for clarity.

🔵 Improvement: apps/isaaclab.python.headless.determinism.kit — Missing documentation comments explaining the key determinism settings
The critical settings for determinism are:

Line 79: omni.replicator.asyncRendering = false
Line 82-84: app.renderer.waitIdle=true, app.hydraEngine.waitIdle=true
Line 165-167: rtx.rtpt.cached.enabled = false, rtx.rtpt.lightcache.cached.enabled = false

These should have comments explaining WHY they enable determinism, not just performance implications. Users looking at this file need to understand what makes it different from the non-deterministic version.

🔵 Improvement: PR description shows command example but no documentation update
The command example in the PR description:

./isaaclab.sh -p ... --experience isaaclab.python.headless.determinism.kit

This should be documented in the official docs (the checklist shows documentation is unchecked). Users won't discover this without documentation.

🟡 Warning: scripts/reinforcement_learning/rl_games/train.py:207 — configure_seed called after environment creation
The configure_seed call happens after gym.make() and wrapper creation (lines 161-189). If environment initialization involves any non-deterministic operations (random initialization, etc.), those won't be affected by this late seeding. The seed should be configured before environment creation, not after. Looking at the code flow:

Line 111: env_cfg.seed = agent_cfg["params"]["seed"] — seed is set in config
Line 161: env = gym.make(...) — environment created
Line 207: configure_seed(...) — deterministic mode enabled too late

This ordering means the environment creation itself may still be non-deterministic even with this fix.

greptile-apps · 2026-04-30T03:23:05Z

Greptile Summary

This PR adds two changes to make Isaac Lab RL training results deterministic for vision-based tasks: a new .kit experience file that enables RTX path-tracing determinism settings, and a configure_seed call in the rl_games train script to enable PyTorch deterministic mode.

P1 — torch.use_deterministic_algorithms(True) unconditionally applied: configure_seed(env_cfg.seed, True) hardcodes torch_deterministic=True for every rl_games training run, not just deterministic/vision ones. This risks RuntimeError on any non-deterministic op and adds overhead without a user-facing opt-in flag. The other RL framework training scripts (rsl_rl, sb3, skrl) are unaffected, creating an inconsistency.
The determinism .kit file is a verbatim copy of isaaclab.python.headless.rendering.kit with only 3 appended lines, which will require manual maintenance to stay in sync.

Confidence Score: 3/5

Not safe to merge as-is; hardcoded torch_deterministic=True is a breaking behavioural change for all rl_games training users.

A P1 finding (unconditional torch.use_deterministic_algorithms(True)) means any rl_games training run using non-deterministic PyTorch ops will fail with a RuntimeError after this change, and all other runs will silently incur determinism overhead. The fix is straightforward, but the current state is not merge-ready.

scripts/reinforcement_learning/rl_games/train.py — the hardcoded torch_deterministic=True argument needs to be made opt-in.

Important Files Changed

Filename	Overview
scripts/reinforcement_learning/rl_games/train.py	Adds `configure_seed(env_cfg.seed, True)` before `runner.load()`, but hardcodes `torch_deterministic=True` for every training run, which could break any rl_games training that uses non-deterministic PyTorch ops.
apps/isaaclab.python.headless.determinism.kit	New kit file for deterministic RTX rendering; effectively a full copy of `isaaclab.python.headless.rendering.kit` plus 3 path-tracing settings. Package title is stale (still says "Camera"), and file lacks a trailing newline.
CONTRIBUTORS.md	Adds author "Xu Li" in alphabetical order — no issues.

Sequence Diagram

sequenceDiagram
    participant User
    participant train.py
    participant gym as gym.make (env)
    participant configure_seed as configure_seed()
    participant runner as rl_games Runner

    User->>train.py: run with --seed 42 --experience determinism.kit
    train.py->>train.py: set env_cfg.seed from agent_cfg["params"]["seed"]
    train.py->>gym: gym.make(task, cfg=env_cfg) → env internally calls configure_seed(seed)
    gym-->>train.py: env created (seed already applied)
    train.py->>configure_seed: configure_seed(env_cfg.seed, True)
    Note over configure_seed: Sets torch.use_deterministic_algorithms(True)<br/>Sets CUBLAS_WORKSPACE_CONFIG<br/>Re-seeds torch/numpy/random/warp
    configure_seed-->>train.py: seed applied (deterministic mode ON)
    train.py->>runner: runner.load(agent_cfg)
    train.py->>runner: runner.run(train=True)
    Note over runner: Any non-deterministic torch op<br/>raises RuntimeError here

_{Reviews (1): Last reviewed commit: "added name to the CONTRIBUTORS.md" | Re-trigger Greptile}

isaaclab-review-bot

🤖 Isaac Lab Review Bot

Summary

This follow-up addresses the previous review findings. The author has made several improvements: added --deterministic CLI flag (addressing the hardcoded True concern), renamed the kit file title to reflect determinism, added documentation to reproducibility.rst, and added explanatory comments for determinism settings in the kit file.

Previous Concerns Addressed

✅ configure_seed now uses args_cli.deterministic instead of hardcoded True
✅ Kit file title updated to "Isaac Lab Python Headless Deterministic Rendering"
✅ Documentation added explaining the --experience flag and --deterministic option
✅ Comments added explaining RTPT cache settings for determinism

Remaining/New Findings

🟡 Warning: scripts/reinforcement_learning/rl_games/train.py:215 — configure_seed still called after environment creation
The previous concern about ordering remains unaddressed. configure_seed() is called at line 215, but gym.make() happens at line 161. Any randomness during environment initialization won't be affected by the deterministic settings. The call should be moved before env = gym.make(...).

🟡 Warning: Test coverage still missing
The PR checklist still shows tests are not added. For a bug fix PR claiming to resolve reproducibility issues (#3505), a regression test demonstrating determinism would strengthen confidence in the fix.

🔵 Improvement: apps/isaaclab.python.headless.determinism.kit — File still missing trailing newline
Minor POSIX compliance issue remains.

Implementation Verdict

Minor fixes needed — The seed ordering issue is the main remaining concern that could affect correctness.

# Description the camera config was importing `isaaclab_physx.renderers` because the default render_cfg was set to that config. this PR sets that to RendererConfig to remove the import, but provides a get_default_render_config method to the backend_utils to lazily import the config if needed. this is called __post_init__ on the camera config to replace the generic config as soon as possible to avoid downstream issues referencing the renderer config. this action can be moved to the factory if downstream references are cleaned up. ## Type of change - Refactor to remove imports in cfg class ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./isaaclab.sh --format` - [x] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] I have updated the changelog and the corresponding version in the extension's `config/extension.toml` file - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there --------- Co-authored-by: nvsekkin <72572910+nvsekkin@users.noreply.github.com>

…ields (isaac-sim#5275) # Description Splits IsaacLab's USD-physics cfg classes into solver-common base classes and backend-specific subclasses, and refactors the writers (`modify_*_properties`, `spawn_rigid_body_material`) so that schema application is data-driven rather than hard-coded per-class. Prepares the schema layer for multi-backend support (PhysX today, Newton/Mjc next) without polluting base classes with silently-ignored fields or stamping backend-specific schemas onto prims that didn't opt in. ## Architecture Two layered concepts: 1. **Per-declaring-class routing.** Each cfg field's USD namespace is determined by the class that declares it (walking the MRO). Base-class fields write under `physics:*`; subclass fields write under their own namespace (`physxRigidBody:*`, etc.). When a `PhysxRigidBodyPropertiesCfg` instance is written, base fields still go under `physics:*` because `_usd_namespace` is read from the declaring class via `__dict__`, not via `getattr` (which would hit the subclass override). 2. **Per-field exceptions.** Some "universal physics" fields have no USD path except through a backend-namespaced attribute today (e.g., `disable_gravity` only exists at `physxRigidBody:disableGravity`). These are declared as `_usd_field_exceptions = {applied_schema: (namespace, [fields...])}` on the base class; the writer applies the exception schema only when one of the listed fields is non-None. The single helper `_apply_namespaced_schemas(prim, cfg, cfg_dict)` in `schemas.py` does both passes for every writer (rigid body, collision, articulation root, joint drive, mesh collision, rigid-body material). ## Design constraints **One cfg class per spawner slot.** Spawners (`UsdFileCfg`, `MeshCuboidCfg`, etc.) carry a single field for each property group: `rigid_props: RigidBodyBaseCfg | None`, `collision_props: CollisionBaseCfg | None`, `joint_drive_props: JointDriveBaseCfg | None`, etc. The user cannot pass two cfgs into the same slot, so the cfg class hierarchy must be **single-rooted per spawner field** — one base class per group, with backend-specific subclasses below. This rules out a "PhysX cfg sits next to a Newton cfg as siblings" design and drives several placement decisions: | Constraint | Consequence | |---|---| | Universal-physics fields must be reachable from any backend's cfg | Goes on the **base** class, not a sibling backend cfg. Users on Newton-only deployments can use `RigidBodyBaseCfg(disable_gravity=True)` without importing `isaaclab_physx`. | | A PhysX-namespaced field whose semantics are universal (e.g., `disable_gravity`) | Lives on the base but routes to the PhysX namespace via `_usd_field_exceptions`. The base stays backend-clean; the writer dispatches the PhysX write only when the field is non-None. | | Writer logic must not branch on cfg subclass | Every writer is the same code path regardless of subclass. The cfg metadata (`_usd_namespace`, `_usd_applied_schema`, `_usd_field_exceptions`) drives behavior; the writer is a pure data interpreter. | | Adding a new backend (Newton, Mjc) | Requires a new subclass with its own `_usd_namespace` / `_usd_applied_schema`. No spawner-side changes, no writer-side changes, no base-cfg-side changes. | | A field has multiple USD paths today (one PhysX-namespaced, one Newton-namespaced) | Belongs on the **PhysX subclass**, not the base. A future `NewtonArticulationRootPropertiesCfg` will own the same conceptual field on the Newton side. ("Rule 2" — e.g., `enabled_self_collisions`.) | | A field has only one USD path today, namespaced under PhysX, but the conceptual quantity is universal | Belongs on the **base** with an `_usd_field_exceptions` entry. ("Rule 1" — e.g., `disable_gravity`, `articulation_enabled`, `contact_offset`, `rest_offset`, `max_joint_velocity`.) When Newton ships its own native attribute, the exception namespace switches transparently with no API change. | ## Field placement ### Base (solver-common) classes — `physics:*` namespace via `UsdPhysics.*API` | Cfg class | Field | USD attribute | |---|---|---| | `RigidBodyBaseCfg` | `rigid_body_enabled` | `physics:rigidBodyEnabled` | | `RigidBodyBaseCfg` | `kinematic_enabled` | `physics:kinematicEnabled` | | `CollisionBaseCfg` | `collision_enabled` | `physics:collisionEnabled` | | `MassPropertiesCfg` | `mass` | `physics:mass` | | `MassPropertiesCfg` | `density` | `physics:density` | | `RigidBodyMaterialBaseCfg` | `static_friction` | `physics:staticFriction` | | `RigidBodyMaterialBaseCfg` | `dynamic_friction` | `physics:dynamicFriction` | | `RigidBodyMaterialBaseCfg` | `restitution` | `physics:restitution` | | `JointDriveBaseCfg` | `drive_type` | `drive:<axis>:physics:type` | | `JointDriveBaseCfg` | `max_force` | `drive:<axis>:physics:maxForce` | | `JointDriveBaseCfg` | `stiffness` | `drive:<axis>:physics:stiffness` | | `JointDriveBaseCfg` | `damping` | `drive:<axis>:physics:damping` | | `MeshCollisionBaseCfg` | `mesh_approximation_name` | `physics:approximation` (token) | | `ArticulationRootBaseCfg` | `fix_root_link` | (synthesizes `UsdPhysics.FixedJoint`) | `JointDriveBaseCfg` and `MeshCollisionBaseCfg` use the typed `UsdPhysics.DriveAPI` / `UsdPhysics.MeshCollisionAPI` accessors at the writer level (multi-instance namespace and `TfToken` with `allowedTokens`, respectively); all other base fields flow through the helper's per-class routing. ### PhysX subclasses — `physx*:*` namespaces, `Physx*API` schemas | Cfg class | `_usd_namespace` | `_usd_applied_schema` | Adds fields | |---|---|---|---| | `PhysxRigidBodyPropertiesCfg` | `physxRigidBody` | `PhysxRigidBodyAPI` | `linear_damping`, `angular_damping`, `max_linear_velocity`, `max_angular_velocity`, `max_depenetration_velocity`, `max_contact_impulse`, `enable_gyroscopic_forces`, `retain_accelerations`, solver iter counts, sleep / stabilization thresholds | | `PhysxCollisionPropertiesCfg` | `physxCollision` | `PhysxCollisionAPI` | `torsional_patch_radius`, `min_torsional_patch_radius` | | `PhysxArticulationRootPropertiesCfg` | `physxArticulation` | `PhysxArticulationAPI` | `enabled_self_collisions`, solver iter counts, sleep / stabilization thresholds | | `PhysxJointDrivePropertiesCfg` | `physxJoint` | `PhysxJointAPI` | (currently empty; reserved for future PhysX-only knobs) | | `PhysxRigidBodyMaterialCfg` | `physxMaterial` | `PhysxMaterialAPI` | `compliant_contact_stiffness`, `compliant_contact_damping`, `friction_combine_mode`, `restitution_combine_mode` | | `PhysxConvexHullPropertiesCfg` | `physxConvexHullCollision` | `PhysxConvexHullCollisionAPI` | `hull_vertex_limit`, `min_thickness` | | `PhysxConvexDecompositionPropertiesCfg` | `physxConvexDecompositionCollision` | `PhysxConvexDecompositionCollisionAPI` | hull / voxel / shrink-wrap tunables | | `PhysxTriangleMeshPropertiesCfg` | `physxTriangleMeshCollision` | `PhysxTriangleMeshCollisionAPI` | `weld_tolerance` | | `PhysxTriangleMeshSimplificationPropertiesCfg` | `physxTriangleMeshSimplificationCollision` | `PhysxTriangleMeshSimplificationCollisionAPI` | `simplification_metric`, `weld_tolerance` | | `PhysxSDFMeshPropertiesCfg` | `physxSDFMeshCollision` | `PhysxSDFMeshCollisionAPI` | `sdf_margin`, `sdf_narrow_band_thickness`, `sdf_resolution`, etc. | ### `_usd_field_exceptions` table These fields are declared on a *base* class but the only USD path today goes through a non-base namespace. Each entry says: "if any listed field on this cfg is non-None, apply the exception schema and write that one attribute under the exception namespace." All other fields on the cfg follow the per-declaring-class routing rule. | Base cfg class | Exception schema | Namespace | Field(s) | Why on the base | |---|---|---|---|---| | `RigidBodyBaseCfg` | `PhysxRigidBodyAPI` | `physxRigidBody` | `disable_gravity` | Per-body gravity exclusion is universal physics; PhysX honors per-body, Newton consumes the same attribute via the bridge resolver (scene-level today; per-body fix is a Newton-side kernel change, not a cfg-API change) | | `CollisionBaseCfg` | `PhysxCollisionAPI` | `physxCollision` | `contact_offset`, `rest_offset` | Collision-pair generation distance and rest gap are universal physics; Newton importer consumes both via PhysX bridge to populate `Model.shape_collision_radius` / `_thickness` (`import_usd.py:2104, 2111`) | | `ArticulationRootBaseCfg` | `PhysxArticulationAPI` | `physxArticulation` | `articulation_enabled` | PhysX honors at sim time; IsaacLab Newton wrapper reads it as a spawn-time guard at `rigid_object.py:1035`. Universal user-facing intent | | `JointDriveBaseCfg` | `PhysxJointAPI` | `physxJoint` | `max_joint_velocity` | Sole USD path to `Model.joint_velocity_limit` in Newton (no `newton:*` equivalent today). The exception namespace switches transparently when Newton ships `newton:maxJointVelocity` as a registered applied API | When any exception field is non-None, the corresponding `Physx*API` schema is applied to the prim. When all exception fields are None, no PhysX schema is stamped — Newton-targeted prims authored from `*BaseCfg` stay free of PhysX schemas they didn't opt in to. ## Field renames (with deprecation aliases) To enforce the convention that python `snake_case` cfg field names map identity-style to USD `camelCase` attribute names, two legacy fields were renamed. Both keep the old name as a deprecation alias forwarded via `__post_init__` (emits `DeprecationWarning`, scheduled for removal in 5.0). | Old name | New name | USD attribute | |---|---|---| | `JointDriveBaseCfg.max_velocity` | `max_joint_velocity` | `physxJoint:maxJointVelocity` | | `JointDriveBaseCfg.max_effort` | `max_force` | `drive:<axis>:physics:maxForce` | ## Type of change - New feature (non-breaking change which adds functionality) - Breaking change (existing functionality will not work without user modification) The split is non-breaking at the spawner-cfg level — every base-class type accepts any subclass via polymorphism, and every legacy `RigidBodyPropertiesCfg` / `JointDrivePropertiesCfg` / `CollisionPropertiesCfg` / `ArticulationRootPropertiesCfg` / `MeshCollisionPropertiesCfg` / `RigidBodyMaterialCfg` / `FixedTendonPropertiesCfg` / `SpatialTendonPropertiesCfg` import path continues to work via deprecation-alias subclasses and `__getattr__` shims on `isaaclab.sim`, `isaaclab.sim.schemas`, and `isaaclab.sim.schemas.schemas_cfg`. Direct attribute access to the renamed fields still works through deprecation aliases. Removal scheduled for 5.0. The breaking aspect: cfg classes in `isaaclab_physx.sim.schemas` and `isaaclab_physx.sim.spawners.materials` are physically relocated. Anyone importing from internal paths (rather than `isaaclab.sim`) needs to update. ## Migration ```python # Before import isaaclab.sim as sim_utils rigid_props = sim_utils.RigidBodyPropertiesCfg(disable_gravity=True, linear_damping=0.1) joint_props = sim_utils.JointDrivePropertiesCfg(max_effort=80.0, max_velocity=5.0) collision_props = sim_utils.CollisionPropertiesCfg(contact_offset=0.02, torsional_patch_radius=1.0) material = sim_utils.RigidBodyMaterialCfg(static_friction=0.7, compliant_contact_stiffness=1000.0) # After (PhysX-targeted) import isaaclab.sim as sim_utils from isaaclab_physx.sim.schemas import ( PhysxRigidBodyPropertiesCfg, PhysxJointDrivePropertiesCfg, PhysxCollisionPropertiesCfg, ) from isaaclab_physx.sim.spawners.materials import PhysxRigidBodyMaterialCfg rigid_props = PhysxRigidBodyPropertiesCfg(disable_gravity=True, linear_damping=0.1) joint_props = PhysxJointDrivePropertiesCfg(max_force=80.0, max_joint_velocity=5.0) collision_props = PhysxCollisionPropertiesCfg(contact_offset=0.02, torsional_patch_radius=1.0) material = PhysxRigidBodyMaterialCfg(static_friction=0.7, compliant_contact_stiffness=1000.0) # After (Newton-targeted — base classes only, no PhysX schemas applied) from isaaclab.sim.schemas import RigidBodyBaseCfg, JointDriveBaseCfg, CollisionBaseCfg from isaaclab.sim.spawners.materials import RigidBodyMaterialBaseCfg rigid_props = RigidBodyBaseCfg(disable_gravity=True) # only base + exception fields available joint_props = JointDriveBaseCfg(max_force=80.0, max_joint_velocity=5.0) material = RigidBodyMaterialBaseCfg(static_friction=0.7) ``` Spawner type annotations remain unchanged — they accept any subclass via polymorphism. ## Internal helper ```python def _apply_namespaced_schemas(prim, cfg, cfg_dict): # 1. Per-field exceptions: pop listed fields, apply exception schema if any non-None, # write under exception namespace. # 2. Per-declaring-class routing: walk MRO to find each remaining field's owner class; # write under that class's _usd_namespace; apply that class's _usd_applied_schema. ``` Used by all five `modify_*_properties` writers and `spawn_rigid_body_material`. Replaced ~125 lines of duplicated gating logic with a single ~30-line helper. ## Side change: configclass `source/isaaclab/isaaclab/utils/configclass.py:_process_mutable_types` now detects string-form `ClassVar` annotations under PEP 563 (`from __future__ import annotations`) so it doesn't wrap `ClassVar[dict]` defaults in `field(default_factory=...)`. Matches Python stdlib `dataclasses` semantics. No pre-existing IsaacLab class used `ClassVar` inside a `@configclass` block, so the change has no effect on existing code; it enables the `ClassVar` metadata pattern this PR introduces. ## Test plan - [x] `test_schemas.py` (38 → 40 tests): all schema-cfg classes write correct attributes under the right namespace; PhysX schemas are NOT applied when only base/UsdPhysics fields are set; deprecation aliases (`max_velocity` → `max_joint_velocity`, `max_effort` → `max_force`) forward correctly and emit `DeprecationWarning`. **40 passed.** - [x] `test_schemas_shim.py`: legacy import paths (`isaaclab.sim.schemas.RigidBodyPropertiesCfg` etc.) resolve via `__getattr__` shims. **All passing.** - [x] `test_articulation.py`, `test_rigid_object_iface.py`, `test_valid_configs.py`, `test_spawn_*` — no regressions. - [x] Full suite (`./isaaclab.sh -t`): 8768/9205 pass, 437 unrelated baseline failures (rendering, `omni.physics.tensors.api` missing, OSC controller, `install_ci`, `pyglet`, Newton env-path, Anymal-C determinism). Zero new regressions; +123 passing tests vs. earlier state. - [x] `./isaaclab.sh -f` (pre-commit) clean. ## Supersedes Together with isaac-sim#5276, supersedes isaac-sim#4847 and isaac-sim#5203 with a cleaner schema-layer design. ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./isaaclab.sh --format` - [x] I have made corresponding changes to the documentation (changelog fragments under `source/isaaclab/changelog.d/`) - [x] My changes generate no new warnings - [x] I have added tests that prove my fix is effective or that my feature works - [x] I have updated the changelog (fragment-based system) and the corresponding version in the extension's `config/extension.toml` file - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there --------- Co-authored-by: ooctipus <zhengyuz@nvidia.com>

…c-sim#5301) Updates docs for using nurec background in locomanipulation sdg ## Type of change - Documentation update ## Checklist - [ ] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./isaaclab.sh --format` - [ ] I have made corresponding changes to the documentation - [ ] My changes generate no new warnings - [ ] I have added tests that prove my fix is effective or that my feature works - [ ] I have updated the changelog and the corresponding version in the extension's `config/extension.toml` file - [ ] I have added my name to the `CONTRIBUTORS.md` or my name already exists there  --------- Co-authored-by: Kelly Guo <kellyg@nvidia.com>

Bumped packages: - isaaclab: 4.6.28 → 4.7.0 - isaaclab_newton: 0.5.26 → 0.6.0 - isaaclab_ov: 0.1.3 → 0.1.4 - isaaclab_ovphysx: 0.1.2 → 0.1.3 - isaaclab_physx: 0.5.29 → 0.6.0 - isaaclab_rl: 0.5.1 → 0.5.2 - isaaclab_tasks: 1.5.34 → 1.5.35 - isaaclab_teleop: 0.3.9 → 0.3.10

## Summary Bumps the Newton pin to [`v1.2.0rc2`](https://pypi.org/project/newton/1.2.0rc2/), which pulls in IsaacLab-relevant fixes plus the upstream tendon-scoping fix. ## What's new in Newton v1.2.0rc2 vs IsaacLab's current pin (`a27277e`) The current IsaacLab Newton pin is from late April; v1.2.0rc2 is the latest release-candidate cut. Notable fixes pulled in: - **[newton-physics/newton#2659](newton-physics/newton#2659 \"Scope USD custom-frequency parsing\" — `parse_usd` now scopes the custom-frequency walk to `root_path` natively. - **[newton-physics/newton#2678](newton-physics/newton#2678 Regression fix. - **[newton-physics/newton#2720](newton-physics/newton#2720 `SolverKamino` reset under `world_mask`. - **[newton-physics/newton#2710](newton-physics/newton#2710 VRAM leak fix on example reset. - Plus 16 other smaller fixes between rc1 and rc2. ## Required dep bumps Newton 1.2.0rc2's \`pyproject.toml\` requires: - \`warp-lang==1.13.0\` - \`mujoco==3.8.0\` (was 3.6.0) - \`mujoco-warp==3.8.0.1\` (was 3.6.0) Pins updated in: | File | Change | |---|---| | \`source/isaaclab/setup.py\` | \`warp-lang==1.12.0\` → \`==1.13.0\`; \`mujoco==3.6.0\` → \`==3.8.0\`; \`mujoco-warp==3.6.0\` → \`==3.8.0.1\` | | \`source/isaaclab_newton/setup.py\` | mujoco / mujoco-warp bumps; Newton pin → \`v1.2.0rc2\` | | \`source/isaaclab_visualizers/setup.py\` | 3× Newton pin → \`v1.2.0rc2\` | | \`tools/wheel_builder/res/python_packages.toml\` | All four pins mirrored | ## Code adapts \`warp-lang\` 1.13 removed the \`wp.math\` namespace. Two IsaacLab call sites use it: - \`source/isaaclab_newton/isaaclab_newton/physics/newton_manager.py:72\` - \`source/isaaclab_ov/isaaclab_ov/renderers/ovrtx_renderer_kernels.py:330\` Both rewritten as \`wp.math.transform_to_matrix(...)\` → \`wp.transform_to_matrix(...)\`. That's the only IsaacLab-side adapt needed. ## Test plan - [x] \`./isaaclab.sh -i newton\` clean install against the bumped pins. - [x] \`pip list\` confirms \`newton 1.2.0rc2\`, \`warp-lang 1.13.0\`, \`mujoco 3.8.0\`, \`mujoco-warp 3.8.0.1\`. - [x] Sanity smoke: Shadow-Hand-Over MAPPO (4 envs, 1 iter) runs clean — simulation init through CUDA graph capture through one training step + checkpoint save, no errors. - [x] Pre-commit clean. ## Caveat Smoke covered Shadow-Hand-Over MAPPO. Other envs with different sensors / renderers / collision setups could surface warp 1.13 or mujoco 3.8 differences the smoke didn't exercise; full PR CI catches them. --------- Co-authored-by: Kelly Guo <kellyg@nvidia.com>

Bumped packages: - isaaclab: 4.7.0 → 4.8.0 - isaaclab_mimic: 1.2.5 → 1.2.6 - isaaclab_newton: 0.6.0 → 0.7.0 - isaaclab_ov: 0.1.4 → 0.1.5 - isaaclab_physx: 0.6.0 → 0.6.1

# Description Clarifies the articulation joint friction API docs across the base, PhysX, and Newton implementations. The base API now warns that joint friction semantics are backend-specific. The PhysX docs distinguish legacy unitless coefficients from PhysX 5 static/dynamic friction efforts and viscous coefficients. The Newton docs now identify joint friction as an absolute force/torque value and include an MJWarp example mapping the value to MuJoCo Warp's `dof_frictionloss`. Fixes isaac-sim/IsaacLab-Internal#875 ## Type of change - Documentation update ## Screenshots Not applicable. ## Checklist - [ ] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./isaaclab.sh --format` - [x] I have made corresponding changes to the documentation - [x] My changes generate no new warnings - [ ] I have added tests that prove my fix is effective or that my feature works (not applicable: docs-only change) - [x] I have added a changelog fragment under `source/<pkg>/changelog.d/` for every touched package (do **not** edit `CHANGELOG.rst` or bump `extension.toml` — CI handles that) - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there

# Description Reduce higher-level dependency on packed state tensors in targeted IsaacLab call sites without changing existing task observation keys. This PR: - changes Pink IK to read `body_link_pose_w` directly instead of slicing `body_link_state_w`; - changes Dexsuite orientation rewards to use `root_link_quat_w` directly instead of slicing `root_state_w`; - adds explicit pick-place helpers for robot link pose and velocity; - keeps `get_all_robot_link_state()` available for compatibility, but marks it deprecated for removal in IsaacLab 4.0; - keeps existing `robot_links_state` task config entries unchanged. Fixes # (issue) ## Type of change - Bug fix (non-breaking change which fixes an issue) - Documentation update ## Screenshots N/A ## Test Plan - `./isaaclab.sh -p -m py_compile source/isaaclab/isaaclab/envs/mdp/__init__.pyi source/isaaclab/isaaclab/envs/mdp/actions/pink_task_space_actions.py source/isaaclab/isaaclab/envs/mdp/observations.py source/isaaclab_tasks/isaaclab_tasks/manager_based/manipulation/pick_place/mdp/__init__.pyi source/isaaclab_tasks/isaaclab_tasks/manager_based/manipulation/pick_place/mdp/observations.py source/isaaclab_tasks/isaaclab_tasks/manager_based/manipulation/dexsuite/mdp/rewards.py` - `./isaaclab.sh -f` - `git diff --check origin/develop..HEAD` - `rg -n "body_link_state_w|root_state_w" source/isaaclab/isaaclab/envs/mdp/actions/pink_task_space_actions.py source/isaaclab_tasks/isaaclab_tasks/manager_based/manipulation/pick_place/mdp/observations.py source/isaaclab_tasks/isaaclab_tasks/manager_based/manipulation/dexsuite/mdp/rewards.py` (no matches) - Existing/new MDP pytest not run locally. Per review, new MDP tests were removed and should be added in a separate PR. Local pytest collection is also blocked in this worktree because `./isaaclab.sh -p` selects `/usr/bin/python3.12` without `torch`. ## Checklist - [x] I have read and understood the [contribution guidelines](https://isaac-sim.github.io/IsaacLab/main/source/refs/contributing.html) - [x] I have run the [`pre-commit` checks](https://pre-commit.com/) with `./isaaclab.sh --format` - [x] I have made corresponding changes to the documentation - [ ] My changes generate no new warnings - [ ] I have added tests that prove my fix is effective or that my feature works - [x] I have updated the changelog and the corresponding version in the extension's `config/extension.toml` file - [x] I have added my name to the `CONTRIBUTORS.md` or my name already exists there Notes: - The unchecked warning item is intentional: this PR adds a `DeprecationWarning` to `get_all_robot_link_state()` so users can migrate before IsaacLab 4.0. - The unchecked test item follows review feedback: MDP tests should be added in a separate PR.

isaaclab-review-bot

Review Summary

This PR adds deterministic rendering settings via a new .kit experience file and wires --deterministic through AppLauncher. The approach is sound — rendering non-determinism from RTPT caches is a real source of irreproducibility in vision-based RL.

Strengths

Clean integration: --deterministic auto-selects the right kit file without user needing to know the filename
Good separation: rendering determinism (kit file) vs PyTorch determinism (configure_seed) are distinct concerns handled independently
Regression test structure is well-designed with proper skip guards for expensive tests
Documentation is thorough with pytest usage examples

Concerns

See inline comments.

redfler added 3 commits April 30, 2026 09:23

add settings to make training results deterministic

8fbf294

checks with ./isaaclab.sh --format

b83b7ee

added name to the CONTRIBUTORS.md

2491117

redfler requested review from Mayankm96, hhansen-bdai, kellyguo11 and ooctipus as code owners April 30, 2026 03:20

github-actions Bot added isaac-sim Related to Isaac Sim team isaac-lab Related to Isaac Lab team labels Apr 30, 2026

isaaclab-review-bot Bot reviewed Apr 30, 2026

View reviewed changes

greptile-apps Bot reviewed Apr 30, 2026

View reviewed changes

Comment thread scripts/reinforcement_learning/rl_games/train.py Outdated

Comment thread apps/isaaclab.python.headless.determinism.kit Outdated

Comment thread apps/isaaclab.python.headless.determinism.kit Outdated

redfler added 3 commits April 30, 2026 08:52

improve title and description, add comment and newline

bef85ea

document --experience and --determinism

980591d

add deterministic argument

5beb625

redfler requested a review from jtigue-bdai as a code owner April 30, 2026 08:54

github-actions Bot added the documentation Improvements or additions to documentation label Apr 30, 2026

isaaclab-review-bot Bot reviewed Apr 30, 2026

View reviewed changes

kellyguo11 reviewed May 1, 2026

View reviewed changes

Comment thread apps/isaaclab.python.headless.determinism.kit Outdated

Comment thread scripts/reinforcement_learning/rl_games/train.py Outdated

Comment thread docs/source/features/reproducibility.rst Outdated

kellyguo11 added this to Isaac Lab May 1, 2026

kellyguo11 moved this to In progress in Isaac Lab May 1, 2026

redfler and others added 9 commits May 7, 2026 01:37

change dependencies

a345583

add flag to AppLauncher

4502b48

add test

33bf92c

update docs

300d494

format

04dba81

Merge branch 'develop' into xul/determinism

fe579e7

isaaclab-bot Bot and others added 5 commits May 8, 2026 07:39

[CI][Auto Version Bump] Compile changelog fragments (schedule)

3fb3e2d

Bumped packages: - isaaclab: 4.7.0 → 4.8.0 - isaaclab_mimic: 1.2.5 → 1.2.6 - isaaclab_newton: 0.6.0 → 0.7.0 - isaaclab_ov: 0.1.4 → 0.1.5 - isaaclab_physx: 0.6.0 → 0.6.1

redfler requested review from pascal-roth and xyao-nv as code owners May 8, 2026 07:39

github-actions Bot added isaac-mimic Related to Isaac Mimic team infrastructure labels May 8, 2026

redfler added 4 commits May 8, 2026 07:44

Merge branch 'develop' into xul/determinism

18197ee

Merge branch 'develop' into xul/determinism

2174553

fix docs

4b60e40

fix docs

9c44458

isaaclab-review-bot Bot reviewed May 9, 2026

View reviewed changes

Comment thread scripts/reinforcement_learning/rl_games/train.py

Comment thread source/isaaclab_tasks/test/test_train_scripts_deterministic.py

Comment thread apps/isaaclab.python.headless.determinism.kit

Comment thread source/isaaclab/isaaclab/app/app_launcher.py Outdated

redfler added 3 commits May 9, 2026 10:57

resolve isaaclab-review-bot

82da515

Merge branch 'develop' into xul/determinism

46fdce9

resolve isaaclab-review-bot

ec23642

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

add settings to make training results deterministic#5449

add settings to make training results deterministic#5449
redfler wants to merge 27 commits intoisaac-sim:developfrom
redfler:xul/determinism

redfler commented Apr 30, 2026 •

edited

Loading

Uh oh!

isaaclab-review-bot Bot left a comment

Uh oh!

greptile-apps Bot commented Apr 30, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

isaaclab-review-bot Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

isaaclab-review-bot Bot left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

Conversation

redfler commented Apr 30, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Description

Type of change

Screenshots

Checklist

Uh oh!

isaaclab-review-bot Bot left a comment

Choose a reason for hiding this comment

🤖 Isaac Lab Review Bot

Summary

Architecture Impact

Implementation Verdict

Test Coverage

CI Status

Findings

Uh oh!

greptile-apps Bot commented Apr 30, 2026

Greptile Summary

Confidence Score: 3/5

Important Files Changed

Sequence Diagram

Uh oh!

Uh oh!

Uh oh!

Uh oh!

isaaclab-review-bot Bot left a comment

Choose a reason for hiding this comment

🤖 Isaac Lab Review Bot

Summary

Previous Concerns Addressed

Remaining/New Findings

Implementation Verdict

Uh oh!

Uh oh!

Uh oh!

Uh oh!

isaaclab-review-bot Bot left a comment

Choose a reason for hiding this comment

Review Summary

Strengths

Concerns

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

7 participants

redfler commented Apr 30, 2026 •

edited

Loading