[ooctipus]

Enable contact-force and joint-wrench rewards with existing NeRD checkpoints—no force labels or extra prediction network. Reconstruction holds applied motor effort fixed and leaves learned motion unchanged.

Opt in with cfg.sim.physics.solver_cfg.reconstruct_forces = True.

Performance — RTX 5090, CUDA graphs enabled:

• Kernel fusion: ~54× faster in the earlier matched projection microbenchmark: 51.81 → 0.95 ms, 128 worlds, FP64, 1,000 iterations.
• Current PR, 4,096 worlds: NeRD step 19.92 → 21.39 ms with reconstruction: +1.47 ms (+7.4%).
• Current PR, 128 worlds: 1.10 → 1.83 ms, adding 0.73 ms.

The kernel comparison measures projection only; current-PR timings include inference, collision and wrench reconstruction. Policy/reward execution and contact-sensor aggregation are excluded.

Accuracy: approximately 16% relative wrench RMS error on 512 held-out walking transitions, with large outliers; 17/512 reached the iteration cap. Remains draft and opt-in; accuracy over learned rollouts is unproven.

Validated: 8 targeted tests, formatting and documentation builds passed. Reconstruction preserves joint positions/velocities bit-for-bit.

Depends on #8825 (and #8822). Review the eight force-reconstruction files. Usage, implementation and limitations.

Contribution 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 fragment and minor-version marker
• Contributor already listed in CONTRIBUTORS.md
• No runtime warnings (existing dependency/backend deprecation warnings remain)