CHL0e/behavior-pi05-baseline

BEHAVIOR-1K turning_on_radio: π0.5 baseline and RAX

Two policies evaluated on the same 10 held-out public test instances (ids 301-310, indices 0-9) of the 2026 BEHAVIOR Challenge task turning_on_radio. These are the 10 instances used for leaderboard reporting. Each instance ran once, with the official timeout of 1.5× the mean human demonstration length (3224 steps at 30 Hz).

Policy	Successes	Mean q-score	Steps on successes
π0.5: the challenge's provided checkpoint	0/10	0.00	-
RAX: program compiled by a Codex agent after 4 exploration episodes and 1 rehearsal	6/10	0.60	344-1136

The q-score is the fraction of BDDL goal predicates satisfied at the end of the episode.

Instance	301	302	303	304	305	306	307	308	309	310
π0.5	✗	✗	✗	✗	✗	✗	✗	✗	✗	✗
RAX	✓	✗	✗	✓	✗	✓	✓	✓	✗	✓

The two runs do not use the same inputs:

	π0.5	RAX
BEHAVIOR-1K	v3.9.3-post2	v3.9.3-post1 (differs only by one robot-config option and CI)
Observations	224×224 RGB + proprioception (DefaultWrapper)	full-resolution RGB-D (720×720 head, 480×480 wrists) + proprioception + odometry
Actions	joint-space actions from the policy	the 23-D native action, plus cuRobo arm motion plans and closed-loop base moves
Trained on	the challenge's teleoperation demos	its own 5 training episodes; the final frames of 2 demos shown as images

On every RAX failure the program ended the episode itself, after 421-718 steps, well before the timeout.

Layout

pi05/turning_on_radio/                  π0.5: per-rollout JSONs, videos, robot config, README
rax/20261020T022131Z_f48626c1b6/         RAX session: README, formal.csv,
    <session>.tar                        transcripts, tool calls, workspace, code and results (2 GB)
    <session>.video.tar                  episode recordings (6.7 GB)

License

CC BY-NC 4.0. The rendered frames show BEHAVIOR-1K assets, which are licensed for non-commercial research use only.