Add the Factory contact-rich assembly task on Newton

[ooctipus]

Description

Important

Depends on #7078 — that PR must be merged first. This branch is stacked on the effort-limit mirror of #6481, so the diff includes its single actuator commit. Review the Factory commits on top; after #7078 lands, rebasing will remove the parent commit from this diff.

Adds IsaacContrib-Factory-Franka, a preset-driven contact-rich assembly task under isaaclab_tasks.contrib.nist.

The task covers 12 NIST taskboard variants:

• M16 nut threading
• round peg insertion at 4, 8, 12, and 16 mm
• rectangular rod insertion at 4, 8, 12, and 16 mm
• small, medium, and large gear meshing

Assembly geometry, reset strategy, curriculum, robot, and physics are composed through presets rather than separate environment classes. The task supports explicit isaacsim_physx and newton_mjwarp presets, with physx retaining automatic backend selection.

Dependency

#7078 separates the two implicit-actuator effort fields:

• effort_limit retains the actuator-facing rated force or torque.
• effort_limit_sim controls the physics-solver clamp.

The Factory Franka configuration intentionally authors distinct rated and solver limits, so it cannot construct on unmodified develop.

Changes

• Added the Factory task, assembly profiles, reset-state generation, sampling, rewards, observations, terminations, and curriculum.
• Added the 12 task presets and RSL-RL configuration.
• Added explicit Isaac Sim PhysX and Newton physics configuration.
• Kept task assets on uploaded asset paths; no Factory-specific Nucleus root constant is introduced.
• Bounded precollected reset states in the existing environment smoke harness so one- and two-environment CI cases remain smoke tests rather than building the full training bank.

Type of change

• New feature (non-breaking)

Validation

• uv run isaaclab -f
• Preset decision suite: 19 passed
• Exact Factory two-environment CI smoke case: 1 passed in 35.8 seconds
• Five-iteration isaacsim_physx training run completed
• Changelog fragment validation
• No new test files were added while addressing the CI failures

[greptile-apps]

Greptile Summary

The PR adds a preset-driven Factory assembly task for Newton and supporting actuator, collision-schema, asset-path, and Newton cloning changes.

• Adds nut-threading and peg-insertion scenes, MDP terms, reset strategies, sampling utilities, curricula, and RSL-RL configuration.
• Preserves authored mesh collision approximations during Newton replication.
• Separates implicit-actuator rated effort limits from solver-level clamps.

Confidence Score: 4/5

The PR should not merge until the accumulator constructs its collision analyzers correctly, because the Factory task otherwise fails during its first reset.

The active Factory reset configuration passes analyzer configuration objects into code that calls their class factory with the predicate-call signature, causing incompatible constructor arguments before any reset states can be collected.

Files Needing Attention: source/isaaclab_tasks/isaaclab_tasks/core/multi_task/utils/event_combinators.py and source/isaaclab_tasks/isaaclab_tasks/core/multi_task/reset_env_cfg.py

Important Files Changed

Filename	Overview
source/isaaclab_tasks/isaaclab_tasks/core/multi_task/utils/event_combinators.py	Adds state accumulation and reset-term composition, but the configured collision acceptance predicates are instantiated with incompatible arguments during precollection.
source/isaaclab_tasks/isaaclab_tasks/core/multi_task/reset_env_cfg.py	Defines Factory reset chains and accumulator configuration, including the CollisionAnalyzerCfg values that reach the broken acceptance invocation.
source/isaaclab_tasks/isaaclab_tasks/core/multi_task/utils/sampling/impl/sampler_warp.py	Adds Warp-backed scoring, probability calculation, and sampling for reset-state selection; no concrete failure was retained.
source/isaaclab_tasks/isaaclab_tasks/core/multi_task/factory_env_cfg.py	Composes the manager-based Factory environment and preset-selected reset and curriculum behavior.
source/isaaclab_newton/isaaclab_newton/cloner/newton_clone_utils.py	Removes blanket post-import convex-hull approximation and delegates authored approximation handling to Newton's importer.
source/isaaclab/isaaclab/actuators/actuator_pd.py	Resolves rated and solver-level implicit-actuator effort limits independently while retaining single-field compatibility.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Factory environment reset] --> B[ACCUMULATOR_RESET]
  B --> C[Run chained scene reset]
  C --> D[Evaluate collision acceptance conditions]
  D --> E[Collect accepted states]
  E --> F[Downsample state table]
  F --> G[Sample reset state per environment]

Loading

_{Reviews (1): Last reviewed commit: "Add skip fragments for the test-only bac..." | Re-trigger Greptile}

[isaaclab-review-bot]

Isaac Lab Review Bot

The Factory task and supporting actuator/collision changes are substantial, but several merge-time issues remain: the task’s asset root is source-checkout-specific, public Newton mesh-simplification APIs are removed without deprecation, the lazy-export stub advertises an undefined symbol, and subset collision checks index packed SDF data incorrectly. There are also a broken type-only import and an avoidable dependency from generic multi-task utilities into the lift task.

• Design and architecture: The preset-driven task structure and moving collision approximation authority from the Newton cloner to asset spawners are coherent. However, the domain-agnostic reset combinators directly depend on core.lift.mdp.events_cfg.SuccessMonitorCfg, coupling shared infrastructure to an unrelated task package. Move that shared configuration into the multi-task utility layer and update both consumers.
• API: ImplicitActuator now preserves independent rated and solver effort limits while retaining the one-field fallback, and CollisionBaseCfg.mesh_collision_property exposes the existing schema dispatch. In contrast, NewtonCfg.simplify_meshes and related callable parameters are removed without the required prior deprecation. LOCAL_ASSET_PATH_DIR constructs a repository-only path that will not resolve in installed distributions, and the mdp lazy-export stub lists an undefined RESET_STRATEGIES symbol.
• Implementation: The SDF collision analyzer packs collider slots and counts using the global environment count and absolute environment IDs, but the query kernel indexes them using the queried subset size and local subset index. Collision validation therefore reads incorrect environments for partial or reordered env_ids; the kernel must use the global stride and absolute ID. The observations.py type-only Offset import also targets a nonexistent mdp.util module and should reference multi_task.utils.pose_offset.

Significant concerns. Posted 6 actionable findings inline.

Automated review; human maintainers own approval decisions.

[AntoineRichard]

AI-generated review

This PR is not ready to merge. The latest head still has functional blockers in task startup and collision validation, and it carries substantial unused/copied infrastructure that should be removed before this becomes a showroom task.

Merge blockers

• The task hard-codes NVIDIA's internal isaac-dev Nucleus, so external users cannot load its assets.
• SuccessMonitorCfg is imported from core/lift/mdp/events_cfg.py, which has been deleted on current develop; the merged tree will fail to import this task.
• Reset precollection calls CollisionAnalyzer with the predicate signature instead of constructing it from its config.
• The SDF kernel addresses globally packed collider data with subset-local strides/indices.
• The process-global collider cache retains stage-owned prims and an old environment count across environment reconstruction.

The prior automated comments about the analyzer constructor, SDF subset indexing, nonexistent .util.pose_offset import, undefined RESET_STRATEGIES, and lift-layer coupling remain valid at this head.

Required PR reduction and consolidation

• Remove/defer the 514-line Warp sampler and the Frontier strategy: no production config in this PR enables warp=True or uses Frontier; only their tests do. Keep the Torch Beta/Uniform path used by Factory.
• Remove the terrain/multi-task command observations and unused reward/termination terms copied into contrib.nist; they are not referenced by the Factory config and create a false public surface.
• Generate the repeated socket/plug asset configs through a small typed helper, consolidate identical keypoint/preset classes, and delete commented-out alternatives.
• Fold the duplicated Hydra import test into the existing shared test and remove config-only smoke/import tests already covered by the shared Factory runtime test and other unit tests. Keep the two scheduler behavior tests in the curriculum test module.
• Add focused regressions for reordered/subset env_ids and for analyzer reconstruction across stages/environment counts. Those tests catch the concrete failures; many current tests only assert trivial properties or implementation strings.

The comments/docstrings also need a full editorial pass: keep them functional and durable, remove author tags/debugging narratives/value-selection postmortems, and delete stale references to terrain, connectors, old layouts, and “backwards compatibility” for new symbols.

I did not run project code, pre-commit, or GPU tests locally because benchmarks are running on this machine. Static git diff --check passed. GitHub CI at this head is currently blocked with failures in two isaaclab_tasks shards, docs, isaaclab_rl, and visualizers.

[ooctipus]

Extracted the implicit-actuator effort-limit semantics into #7078, the effort-limit mirror of #6481, and rebased this branch on that single prerequisite commit. The direct diff from #7078 to this head now touches only isaaclab_tasks.

Revalidated after the split:

• preset decision suite: 19 passed
• exact IsaacContrib-Factory-Franka-2-cuda smoke case: 1 passed in 35.8 seconds
• repository formatter and changelog checks: passed