Activity
A dated log of what I have been working on, newest first. Open a day to see the write-ups from it.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
BEHAVIOR 2 entries
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
- Recovery with cuRobo: v6c, development and held-out With a torso-up retry for the search posture and the arm tuck planned by cuRobo, v6c gives a clean recording on 4 of 10 never-seen tasks, up from 3, and on 2 of 5 development cases, up from 1. setting_the_fire 205 and setting_mousetraps 202 are new clean recordings.
- World action models compared for BEHAVIOR-1K Of six video-based world action models, only Flex-pi both runs and fine-tunes on a 24 GB GPU under a clean license, and none has a BEHAVIOR-1K result yet.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
BEHAVIOR 3 entries
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
- Search v2 replays Six versions of the 4 recovery cases side by side, from the broken v2 to the A, B and C experiments, where C passes 2 of 4 without knocking the table.
- Active search, before and after The not-visible recovery cases rerun with a search that drives to spots and sweeps the head and wrist cameras, old and new attempt side by side.
- Why the scripted recoveries fail All 110 failed recovery attempts on the pilot cases, plus the bookcase chains and the new-task failures, grouped by cause with a clip and a feedback box for each group.
BEHAVIOR 4 entries
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
- Recovery pilot results so far Scripted recoverers on pt50's near-miss failures produce 19 confirmed recordings in 132 attempts, and the stall fix turns the 27 pick timeouts into named failures but adds only 2 recoveries.
- All 19 confirmed recoveries A gallery of every detector-confirmed clip from the recovery pilot, grouped by the 8 pt50 failure cases they come from, with seeds passed per case.
- Why the scripted recoverer times out on picks In the issue 74 recovery pilot, pick attempts hit the 40 s cap with the hand stuck 10 to 35 cm from the object, because the recoverer never moves the base again once it starts reaching.
- Replay of a pt50 failure, assembling_gift_baskets 200 The first sim viewer replay, with all three cameras, the Q plot and the five spans where Q stays flat on the way to a 0.69 failure.
BEHAVIOR, OPENVLA 6 entries
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
- The 99-task submission, task by task pt50 scores 0.155 with 18 successes on the original 50 tasks, while the step follower scores 0.031 with none on the new 49.
- Where pt50 gets stuck Only 12 of 453 failures lock against a joint limit. Parking the base while the arms keep working is more common and has several causes.
- Planner and follower, where things stand pt50 hardly listens to its prompt, and with a perfect oracle planner the team's follower still gains only about half of what pt50 does per phase.
- SnapFlow one-step student in sim Every step-reduced arm loses success rate against pt50 at 10 steps and saves almost no time, so SnapFlow is parked.
OPENVLA
Does fine-tuned OpenVLA use the instruction? Language ablations on LIBERO.
- OpenVLA language ablation on LIBERO-Goal: results Across 800 paired episodes the original instruction succeeds 170 of 200 times (85%), blank, missing and wrong instructions all go 0 of 200, and a wrong instruction makes the robot do the commanded task in 148 of 200.
- OpenVLA language ablation on LIBERO-Goal: pilot Pilot on DeltaAI GH200: the original instruction succeeds on 25 of 30 rollouts (83%), and blank, missing or wrong instructions each go 0 of 3 on task 0.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
BEHAVIOR 1 entry
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
SWOT, BEHAVIOR 5 entries
SWOT
Internal-wave detection in SWOT imagery.
BEHAVIOR
BEHAVIOR-1K 2026 challenge work.
- turning_on_radio before and after the evaluator fix A stale step-0 camera frame in the batched evaluator, the fix, and every test instance across five runs.
- bringing_in_wood crash, instances 309 and 310 PhysX returned NaN for a gripper finger pose about 5 minutes into the warm-up rerun.
- cook_hot_dogs before and after the warm-up fix Task Q fell from 0.30 to 0.05, and the evidence fits seed noise better than a real regression.
- pt50 warm-up rerun: biggest gains and drops The four tasks that moved up the most and the four that moved down, with side-by-side timelapses.