[ooctipus]

Description

Add native offline NeRD workflows for Cartpole and flat/rough ANYmal C: collect transitions through recorder terms, fit resumable dynamics from HDF5, and train or play policies through Isaac Lab's normal reinforcement-learning entrypoints.

NeRD consumes the actual applied motor effort at each physics step. Actuators retain position/velocity commands, clipping and internal state. The Newton solver owns learned physics and CUDA-graph state writes; task configurations own capture schemas and recipes; isaaclab_rl.nerd owns fitting. Rough terrain preserves live contact conditioning, terrain-aware resets and checkpoint validation.

Dependency: #8822 provides shared capture and optional batched episode storage. This branch includes that unmerged commit and targets upstream main from the fork. Review only the NeRD changes. NeRD is independent of the separate SuperMimic/G1 PR #8826.

The NeRD guide covers recording, fitting, policy training, logging, checkpoint upload and recovery. Newton solver replacement derives its manager class and removes the corresponding task workarounds. Legacy target-conditioned/incomplete checkpoint contracts remain rejected; no compatibility aliases are added.

Validation

The full warnings-as-errors Sphinx HTML build passed for this branch.

• In this isolated branch, 5 NeRD tests passed: four capture/fit/resume/deployment configurations plus the relocated NeRD architecture gate. The cases cover Cartpole/ParallelNeRD, flat ANYmal/ParallelNeRD, flat ANYmal/VectorNeRD and rough ANYmal/ParallelNeRD, including applied-effort capture, contact history, checkpoint mismatch rejection and repeated graph resets.
• Newton solver-configuration regressions: 8 passed in a separate process.
• Full repository formatting/changelog gates and git diff --check passed.
• Verified imports resolve to this checkout with installed editable-package fallback disabled; the SuperMimic modules are unavailable. The runtime implementation matches Integrate native NeRD and SuperMimic with Cartpole, ANYmal C and G1 #8806 byte-for-byte. Existing ownership assertions moved from the SuperMimic test into the existing NeRD test file; no production abstraction was added for the split.
• Fresh fitting and policy integration results from Integrate native NeRD and SuperMimic with Cartpole, ANYmal C and G1 #8806 remain applicable: ANYmal fitting, Cartpole fitting, recovery and PPO smokes, and rough-terrain training. Splitting this PR does not constitute a new training run.

Dependencies and limits

PhysicsNeMo Robotics is optional. The validated environment uses development revision a81160ef and local attention-accumulation fix ee6d2e0, with PhysicsNeMo 2.1.1. Push access to that dependency was denied; safe FP16 fitting still depends on the unpublished fix, and a released compatible package combination remains unresolved.

The solver fixes robot parameters/materials while permitting supported static terrain changes through live contacts. Independent moving objects, per-surface materials, external body wrenches and force/joint-wrench sensors remain unsupported. Solver-side motor drives and MuJoCo-specific actuators are rejected when they generate effort beyond the captured boundary. Existing wrench diagnostics and optimized reconstruction kernels remain external experiments; this PR does not advertise sensor accuracy or learned-rollout convergence.

Type of change

• Native learned-physics integration, task configurations and offline fitting
• Solver-configuration fixes and checkpoint-contract migration
• Documentation update

Checklist

• Read the contribution guidelines
• Ran repository formatting and changelog checks
• Updated documentation
• Validated changed behavior and applied the test-audit criteria
• Added package changelog fragments and independent minor-version markers
• Contributor already listed in CONTRIBUTORS.md
• No runtime warnings (existing dependency/backend warnings remain)