Hoshipu/reflex-b1k-task1-2d

ReFLEX · BEHAVIOR-1K task 1 (picking_up_trash) · failure modes + mode planners (ablation arm 2d)

This is ablation arm 2d of the ReFLEX project: the failure-mode pipeline and its recovery planners, with no failure multiplication. It is a companion to arms 2b (Hoshipu/reflex-b1k-task1-recoveries) and 2c (Hoshipu/reflex-b1k-task1-2c). It uses the same task, policy (pi0.5 2025ckpt2), stack and file layout. See the 2b card for the layout and the conversion notes.

How it was made

1. Failure modes: the five modes of the full ReFLEX round trash-r1 (modes.json in fastwalker1118/embodiedclaw_data, failures/task-0001/2025ckpt2/trash-r1).
2. Detection over all 499 rollouts: trash-r1's frame-picking code is not published, so an agent re-implemented its frame rules from modes.json and the trash-r1 report. It reads the rollouts' saved states offline. On trash-r1's own 268 frames (43 episodes) it recalls 255 (95%), between 0.91 and 0.98 per mode. The sweep found 2,738 frames in 354 rollouts (detected_frames.json, DETECTORS.md). The 20 public test instances are excluded.
3. Recovery: each mode runs trash-r1's frozen planner (EmbodiedClaw benchmarks/b1k/tasks/01_picking_up_trash; recovery.toml gives the versions and success rules). The planners run unchanged, behind a thin adapter for OmniGibson 3.9.2. Frames were drawn in turn from the ten (mode, kind) groups, preferring new episodes, and each frame ran once with no variants. A run counts by recovery.toml's rule. 261 runs were made and 246 recovered (94%). Sampling stopped at about 2.5 h of recovery data, to match arms 2b and 2c.

Contents

Clips	246, from 161 rollouts
Frames	273,530 at 30 fps, 2.53 h (26 % of the 200 task-1 human demos)
Whole task finished within the clip	76
Share of frames by mode	last_can_out_of_view 38.8 %, lying_can_grasp_misses 19.0 %, can_released_outside_ashcan 17.1 %, can_grasped_before_ashcan 13.4 %, ashcan_grasped_left_hand 11.8 %

In scored.json, planner is 2d-frozen/<mode>/<kind>/<run>, and goal_fact is the fact that made the run count. Paths are relative (root: "."). The parquets use the 2025 layout: actions are one row late and the base velocity is in the world frame. Convert them with mix_dataset.py before training. Videos are RGB only.