rby1-fold-v1.1

Rainbow Robotics RB-Y1 T-shirt folding — DAgger correction set (2026-06-29) with per-frame advantage labels, LeRobot v3.0. Human-in-the-loop intervention rollouts collected on top of the rby1-fold-v1.0 policy, then hand-edited and recovered.

• 40 episodes / 81,813 frames @ 30 fps (~45.5 min)
• task (single, per-frame task_index): fold the t-shirt (0)
• cameras: cam_high, cam_left_wrist, cam_right_wrist (480×640, AV1)
• state / action: 26-dim — torso(6) + right_arm(7) + left_arm(7) + head(2) + wheel(2) + right_gripper(1) + left_gripper(1)
• every frame flagged observation.is_intervention = true (expert-correction segments)

Relation to v1.0: v1.0 is the cleaned imitation master (126 ep). v1.1 is the next round of on-policy corrections — the human takeovers gathered while running the v1.0-trained policy, curated with dataset_segment_editor and re-encoded. Intended to be merged into / co-trained with v1.0 for AWBC-style fine-tuning.

Columns

observation.state, action, observation.images.*, observation.is_intervention, timestamp, frame_index, episode_index, index, task_index.

Labels (annotations/)

Advantage labels predicted by the RB-Y1 stage-advantage estimator (rby1_advantage_estimator_113_260627/step_010000), horizon 50, computed per frame (image-space, resolution-independent — aligned 1:1 with the frames here).

• eval_progress.json — {episode_index: [stage_value_pred, ...]}, the estimator's per-frame stage-progress value V(s).
• advantage.json — per-frame relative_advantage = V(s+50) − V(s) and absolute_advantage. For AWBC prompting, the global p70 threshold of relative_advantage is 0.032 (frames above → Advantage: positive).

Note

meta/stats.json numeric stats are computed from state/action; run compute_norm_stats before training. Videos are AV1 at native 480×640.

Changelog

• v1.1 (2026-06-29): DAgger correction set — 40 curated/recovered intervention episodes collected against the v1.0 policy, plus estimator advantage labels.
• see rby1-fold-v1.0 for the base imitation master and its lineage.