dgrachev/droid_1.0.1

DROID 1.0.1 for pi05 training: successful, non-idle segments

lerobot/droid_1.0.1 rewritten twice, videos untouched. First augment_droid_delta_ee_gripper_events.py (lerobot_policy_framepick) added action.delta_ee, observation.extrinsics.static1/static2/wrist1 and the observation.gripper.time_* columns; the flat action is the delta-EE command and the flat observation.state the EE state. Then benchmarks/robolab/augment_droid_dataset.py (this revision) mirrored the data pipeline of openpi's pi05_full_droid_finetune:

• only successful episodes are kept (78864 of 95658);
• every episode is cut into its non-idle runs (openpi's compute_droid_nonidle_ranges.py rule on action.joint_velocity: idle runs of 7+ frames removed, runs of 16+ frames kept), each run its own episode: 106920 episodes, 22129092 frames, 93.3% of all frames kept (93.3% in openpi's published ranges);
• action.joint and observation.state.joint hold [joint_position(7), gripper_position(1)];
• observation.state.ee_<rotation>, action.ee_absolute.<rotation> and action.ee_delta.<rotation> hold the EE state, the absolute EE command and the per-step displacement it asks for, each as [xyz, rotation, gripper] for rotation in rotvec, quat_xyzw, rot6d, rot6d_rows, matrix, euler_xyz_fixed: the layout of dgrachev/libero_extrinsics. action.ee_delta.rotvec equals the first hop's action.delta_ee;
• every names entry is a flat list.

meta/stats.json holds global histogram quantiles over all frames. The action.joint entry describes 16-step chunks minus the state at chunk start (gripper absolute), the values pi05 normalizes under use_relative_actions; re-aggregating it from the per-episode statistics would replace it with absolute values, so rerun the script instead. Per-episode image statistics are those of the parent episode.

The per-offset tables framepick normalizes with are stats keys, not columns, shaped (16, action_dim); row i describes offset i inside the chunk, so a shorter chunk keeps rows :chunk_size. For every action.ee_absolute.<rotation> column: .anchored_horizon_16 (chunks minus the state at chunk start, gripper absolute), .expressed_<frame>_horizon_16 for base, static1, wrist1 (the absolute poses as seen from the frame, frozen at chunk start) and .expressed_horizon_16 pooled over those frames, .anchored_expressed_horizon_16 (anchored, then the frame's basis change, pooled). Every 10-th complete chunk inside one episode contributes: 2052569 chunks.