ghkim-rlwrld/shell-game-pi05
0
shell_game pi05 checkpoints
Three pi05 policies trained on the shell_game real-robot dataset (45 episodes, 7-DoF single arm plus gripper, stereo x3 cameras, 9.26 Hz measured control rate).
rot6d rather than rpy: roll wraps at +-pi and produced 269 discontinuities across the 45 episodes; unwrapping and converting to rot6d leaves 0 (std 3.08 -> 0.039).
Each directory holds params/ and assets/ for the final step. train_state/ (optimizer moments, 31 GB per step) is deliberately not uploaded -- these load for inference, not to resume training. assets/ carries the normalisation statistics and is required to get correct actions.
No evaluation has been run against these checkpoints yet.