active-adaptation

Implement and wire MDP terms in active-adaptation (observations, rewards, terminations, actions, commands, randomizations) using the V2 deferred-init API. Use when adding or modifying classes under envs/mdp/, editing cfg/task/ observation/reward/termination/input/command/randomization blocks, or debugging env_base step/reset callbacks.

Implement, extend, and run off-policy RL algorithms in active-adaptation (SAC, distributional SAC, SimbaV2, RLPD, BAC/BEE). Use when adding or modifying algo files under learning/offpolicy, wiring Hydra algo configs, debugging train_offpolicy.py runs, replay buffers, or critic/actor training loops.

Implement, extend, and run on-policy RL algorithms in active-adaptation (PPO, symmetry augmentation, SPO, Muon). Use when adding or modifying algo files under learning/ppo, wiring Hydra algo configs, debugging train_ppo.py runs, GAE/advantage computation, or trust-region policy updates.

Read WandB run diagnostics (grad norms, losses, PPO KL/entropy, explained variance) to analyze numerical stability and training behavior for RL algorithms in this repo. Use when the user provides a WandB run URL or run id and asks for algorithm analysis from logs.