--- tags: [LeRobot, so101, real-robot, block-stacking, task-decomposition] configs: [] --- # SO-101 block stacking: phase-labelled demos The same 98 episodes as `hrl-imerse/doug-stack-98-trimmed`, with one added feature: `observation.environment_state` (float32[6]) is a one-hot encoding of the current sub-task: `pick_orange, place_orange, pick_green, place_green, pick_purple, place_purple`. **How the labels are made:** the phase comes only from the measured gripper position (`observation.state[5]`). - Holding a block reads about 8.7, open reads 27-44, and closed on nothing reads below 6. - Pick switches to place once the gripper has stayed in the 6-12 band for 6 frames. - Place switches to the next pick once the reading has stayed above 16 for 6 frames. The same rule runs live on the robot, so labels at training time and test time match. **Excluded episodes:** 84 episodes contain exactly 3 grasps. The other 14 contain regrasps, so their labels are ambiguous; they were excluded from training with `--dataset.exclude_episodes=[2,18,23,29,33,40,44,52,66,75,78,81,90,97]`. The model `hrl-imerse/doug-act-stack-phases-15k` was trained on this data.