Where pt50 gets stuck

BEHAVIOR-1K 2026 challenge, pt50 submission run on the original 50 tasks, 468 of 500 episodes

The question was whether pt50 fails because it gets stuck against its own joint limits. The submission run logged joint positions and commanded targets at every 30 Hz sim step, and the eval logs map each env to its instance, so every matched episode can be checked. Locking against a joint limit is real but rare. The more common pattern is the base parking while the arms keep working, and the videos show several different causes behind it.

Locked at a joint limit

A joint counts as pushing its limit when it is within 0.05 rad of its URDF limit and pt50 still commands it further out. An episode is locked to the end when one joint does this for most of every minute, at least 2 minutes in a row, through the last minute.

12 of 453 failures end locked, and none of the 15 successes do. 10 of the 12 are the left wrist roll (left_arm_joint7) at +1.571 rad, and the other two are wrist joints (left_arm_joint6 and right_arm_joint6). The lock usually starts about 70% of the way into the episode and lasts until timeout. A guard that only targets joint limits would rescue at most about 3% of failures.

Tasks with at least one failure locked at a limit until timeout

task	episodes	failures	failures locked to the end	main joint
clean_a_patio	10	10	2	left_arm_joint7
getting_organized_for_work	10	10	2	left_arm_joint7
moving_boxes_to_storage	8	6	2	right_arm_joint6
tidying_bedroom	6	6	2	left_arm_joint7
clean_up_your_desk	10	10	1	left_arm_joint7
rearranging_kitchen_furniture	10	10	1	left_arm_joint6
setting_the_fire	10	10	1	left_arm_joint7
sorting_household_items	10	10	1	left_arm_joint7

Base parked, arms busy

A minute is flagged when the base is nearly still (speed under 0.03, the slowest 10% of success minutes) while the arms move more than in 90% of success minutes. Flagged minutes are 8.3% of failure minutes against 1.8% of success minutes, and they build up over the episode, from 3.5% in the first third of failed episodes to 10.5% and 11.8%. 59 of 453 failures spend at least a quarter of their minutes this way. The control is weak, since 15 successes give only 56 success minutes.

Most near-miss tasks barely show it. putting_shoes_on_rack is at 0%, setting_mousetraps 2% and picking_up_toys 5%, while sorting_vegetables is at 46%.

Most flagged tasks

task	share of failed-episode minutes flagged	failures
sorting_vegetables	46%	10
bringing_in_wood	38%	6
turning_on_radio	29%	7
clean_a_patio	21%	10
chop_an_onion	20%	10
tidying_bedroom	19%	6
collecting_childrens_toys	15%	10
cook_bacon	15%	10
loading_the_car	13%	9
storing_food	12%	10
bringing_water	12%	6
make_pizza	12%	4

What the flagged runs look like

Each sheet shows six moments of one failed run, with the head camera on the right and the two wrist cameras on the left of each tile.

Six frames of sorting_vegetables instance 310. From 2.5 minutes on the robot holds the vegetable basket up with both arms. — sorting_vegetables 310. The robot lifts the whole vegetable basket and waves it around for 6 minutes without placing anything.

Six frames of clean_a_patio instance 302. After 4 minutes the cameras point at the sky and the house. — clean_a_patio 302. The robot is lost, with the head and wrist cameras pointing at the sky and the house and nothing to work on in view.

Six frames of moving_boxes_to_storage instance 307. From 4.7 minutes the robot carries the box tilted up. — moving_boxes_to_storage 307. The robot carries the box tilted up from minute 4.7 to the end, its right wrist pinned at -1.047 rad while pt50 commands -1.26.

Takeaway

Getting stuck is a real failure mode, but joint limits explain only a small part of it. Most of it is pt50 looping in a state where it does not know the next step, holding the wrong thing or lost. Knocked-over objects cannot be seen in joint data, because object state is not logged, and none of the three runs above shows one. Since the causes differ, a single scripted fix is unlikely to cover them, which points back to recovery data on these tasks.