Brankozz/pi05_g1_stack_cube

pi0.5 fine-tuned on Unitree G1 stack-cube (action-horizon A/B)

Fine-tuned from pi05_base (openpi) on dataset yigao7117/stack_cube_v3 (Unitree G1 + Dex1 grippers, 16-dim joint state/action, 3 cameras, 30 fps, 50 episodes).

• h30/ — action_horizon=30 (1.0 s chunk), steps 5000–29999
• h15/ — action_horizon=15 (0.5 s chunk), steps 5000–29999

Each step dir contains params/ (EMA weights, for inference) and assets/ (norm stats). Delta joint actions (14 arm dims), absolute grippers; quantile normalization. Train configs: pi05_g1_stack_h30 / pi05_g1_stack_h15 in the accompanying openpi fork.