assembling_gift_baskets 202, pick pillar_candle_86
left armpick3 of 3 seedsThe phase is to pick up the pillar candle from the coffee table.
BEHAVIOR-1K 2026 challenge, issue 74 recovery pilot, gallery as of Oct 22
These are all 19 recoveries that the phase detector confirmed in the issue 74 recovery pilot. Each one starts from a pt50 failure state in batch_a, restored in the sim, and a scripted recoverer with privileged object state finishes the next pick or place. The 19 come from only 8 distinct failure cases. Seeds 0 to 2 only add a little position and rotation noise to the same start, so three passing seeds of one case are close to the same recovery three times. That is why the pilot is scaling to 75 more near-miss cases and adding human takeovers recorded from the sim page.
For the counts across all 44 cases see recovery pilot results so far, and for why picks used to stall see why the scripted recoverer times out on picks. The work is tracked in issue 74.
One row per failure case. Seeds counts the seeds with at least one confirmed attempt. Attempts counts confirmed attempts over all attempts, so the toy train case 201 shows its two reruns.
| case | phase | seeds passed | attempts confirmed | run |
|---|---|---|---|---|
| assembling_gift_baskets 202 | pick up the pillar candle from the coffee table | 3 of 3 | 3 of 3 | first pass |
| assembling_gift_baskets 200 | place the pillar candle in the wicker basket | 2 of 3 | 2 of 3 | first pass |
| collecting_childrens_toys 203 | pick up the dice from the bed | 3 of 3 | 3 of 3 | first pass |
| collecting_childrens_toys 204 alt | pick up the toy train from the desk | 3 of 3 | 3 of 3 | first pass |
| collecting_childrens_toys 201 alt | pick up the toy train from the desk | 3 of 3 | 3 of 5 | first pass and stall-fix rerun |
| preparing_lunch_box 205 | pick up the half apple from the chopping board | 3 of 3 | 3 of 3 | first pass |
| putting_up_Christmas_decorations_inside 206 | pick up the pillar candle from the sofa | 1 of 3 | 1 of 3 | first pass |
| putting_up_Christmas_decorations_inside 207 | pick up the pillar candle from the sofa | 1 of 3 | 1 of 3 | first pass |
| 8 cases | 19 of 24 | 19 of 26 |
Each clip is the full attempt in real time, from the restore to the end of the phase. The panels are, from left to right, the left wrist camera, the right wrist camera and the head camera. Clips load only when played.
The phase is to pick up the pillar candle from the coffee table.
The phase is to place the pillar candle in the wicker basket.
The phase is to pick up the dice from the bed.
The phase is to pick up the toy train from the desk. This is the alternative case the labeller made from the same failure.
The phase is to pick up the toy train from the desk. This is the alternative case the labeller made from the same failure. Seeds 0 and 2 timed out on the first pass and were confirmed in the stall-fix rerun.
The phase is to pick up the half apple from the chopping board.
The phase is to pick up the pillar candle from the sofa.
The phase is to pick up the pillar candle from the sofa.
Clips and counts come from the attempt.mp4 and result.json of every attempt under recov50/runs/pilot and recov50/runs/rerun_plateau. The smoke runs are left out.