W&B and TensorBoard

Training configs default to TensorBoard through training.logger=tensorboard. Set training.logger=wandb to use Weights & Biases. The shared W&B fields live in the training config blocks, including wandb_project, wandb_entity, wandb_group, wandb_name, wandb_tags, wandb_notes, and wandb_mode.

Examples

uv run train --algo ppo --task go2_joystick_flat --sim mujoco training.logger=tensorboard

uv run train --algo ppo --task go2_joystick_flat --sim mujoco \
  training.logger=wandb \
  training.wandb_project=unilab

ExperimentTracker in src/unilab/training/experiment.py writes run_config.json and run_summary.json in the run directory. For RSL-RL PPO, the script also patches the RSL-RL W&B writer when training.logger=wandb.