Language

NaN Visualizer¶

PPO has a NaN guard under training.nan_guard in src/unilab/conf/ppo/config.yaml, enabled by default to match APPO and off-policy. When active, src/unilab/scripts/train_rsl_rl.py installs NanGuard, checks observation dicts and rewards, and writes a .npz dump plus model metadata when it detects NaN/Inf values.

uv run train --algo ppo --task go2_joystick_flat --sim mujoco \
  training.nan_guard.enabled=true \
  training.nan_guard.output_dir=/tmp/unilab/nan_dumps

The viewer implementation is src/unilab/utils/nan_viz.py, registered as the unilab-viz-nan console entry. It replays a dump path and lets you select the environment index. Dump format and round-trip loading are covered by tests/test_nan_guard.py.