Hoshipu/reflex-b1k-task1-2c

ReFLEX · BEHAVIOR-1K task 1 (picking_up_trash) · recoveries + failure multiplication (ablation arm 2c)

Ablation arm 2c of the ReFLEX project: per-failure agent recovery plus failure multiplication, without failure modes. Companion of arm 2b, Hoshipu/reflex-b1k-task1-recoveries. The task, policy (pi0.5 2025ckpt2), stack and file layout are the same as there; read that card for the layout and the conversion notes.

How it was made

1. Seeds: 50 failures drawn at random (seed 2026) from arm 2b's 347 recorded recoveries. In 2b, one Claude Code agent per failed rollout found the failure from video, restored a restart frame and wrote a privileged classical planner for the failed subtask. No agent saw failure modes.
2. Perturbation: one generator agent (claude-opus-5-5, $2.09) received each seed's failure description and start state and wrote 5 accepted variants per seed (J = 5). A variant displaces the robot base (±8–30 cm, ±0.12–0.35 rad), the arm and torso joints, and free floor cans (≤10 cm) of the restart frame. The simulator settles each variant for 0.5 s and rejects it when an object drifts, a joint pose cannot be held, a grasp changes, the robot touches something new, or an Inside or goal fact changes. 386 candidates were tried and 347 accepted (90%). See VARIABLES.md; variants.json lists the edits.
3. Recovery: each variant ran its seed's frozen 2b planner (no further agent iteration), recorded directly. 233 of 250 gained the seed's goal fact (p = 93%) and are kept.

Contents

Clips	283: the 50 seeds' 2b recoveries (E = 1) and 233 recovered variants (E = 1 + NN for variant vNN)
Frames	270,117 at 30 fps = 2.50 h (25.6 % of the 200 task-1 human demos, 1,053,550 frames)
Whole task finished within the clip	122
Recovered subtask	195 Inside (can into bin), 88 IsGrasping (pick)
Token cost	$102.67 ($100.58 for the 50 seeds' 2b agents + $2.09 for the generator)

The episode id is task_id*10000 + instance_id*10 + E. In scored.json, planner says which clip is which: 2b-agent/epNNNNN for a seed, 2c-variant/epNNNNN/vNN for a variant. Paths are relative (root: "."), so run mix_dataset.py from the dataset root. As in 2b, the parquets use the 2025 layout (actions one row late, base velocity in the world frame). Convert them through mix_dataset.py before training. All videos are RGB only.