[ooctipus]

Adds native G1 SuperMimic training through Isaac Lab's standard RL scripts, with MLP or optional one-frame NeRD dynamics, differentiable policy learning, evaluation and video. The task owns observations, objectives and transition capture; the learner owns replay and optimization. Checkpoints include optimizer state and support W&B uploads.

Joint motion and velocity control: set env.commands.motion.velocity.fraction=0.5. The interfaces share a world model and action decoder, with separate encoders and critics. Motion tracking updates the decoder; velocity return updates only its encoder through the frozen decoder. Reference-command rollouts add latent alignment. Captured command provenance preserves real/imagined command timing. Motion-only training remains the default.

Dependencies: #8822 supplies shared capture/storage; this branch includes its unmerged implementation. #8825 is independent. The optional NeRD network requires PhysicsNeMo Robotics; the separate normalization integration remains in #8835.

Usage guide covers data preparation, both training interfaces, logging, resume, motion evaluation and held-velocity-command playback.

Validation:

• 94 focused learner, G1 task and entrypoint tests passed. Removing decoder freezing makes the gradient-routing regression fail.
• Native joint MLP and NeRD smoke training passed with CUDA graphs enabled; MLP checkpoint resume, motion-only training, prior checkpoint-schema resume and both playback modes passed.
• Full formatting/changelog checks and warnings-as-errors Sphinx HTML build passed.

These are integration checks using short training runs. They do not establish velocity-policy convergence, an exact paper reproduction, or the paper's zero-failure result. Robot and motion data remain separately prepared inputs.