Training and Playback

PACE system identification is the primary workflow; the RSL-RL entry points remain for existing project templates. The bundled task has a zero-weight reward, so it is an environment/API compatibility check rather than a useful locomotion-training benchmark until you add a task reward.

Training is local by default and uses TensorBoard (no W&B login or upload):

uv run python scripts/rsl_rl/train.py \
  --task Isaac-Pace-Anymal-D-v0 --num_envs 64 --max_iterations 1 \
  --logger tensorboard

Historical flags including --num_envs, --max_iterations, --seed, --logger, --checkpoint, --device, --headless, and boolean --video are translated by the wrappers. The native mjlab form is shown by --help. Cloud logging is opt-in with --logger wandb; add wandb/ to local ignores if you enable it.

play automatically chooses the newest local model_*.pt under logs/rsl_rl when no checkpoint is supplied. You can be explicit instead:

uv run python scripts/rsl_rl/play.py \
  --task Isaac-Pace-Anymal-D-v0 --checkpoint logs/rsl_rl/pace_sim2real/<run>/model_1.pt

For a no-checkpoint environment smoke test use --agent zero or --agent random.