jermo1120/b1k-spraying-for-bugs-pi05-eval

BEHAVIOR Challenge 2026: task 38 spraying_for_bugs evaluation

Evaluation of jermo1120/b1k-spraying-for-bugs-pi05 (checkpoint step 49999) with the official BEHAVIOR-1K v3.9.3 evaluator: public test instances 0–9, 1 rollout each, videos enabled.

Pick up the pesticide atomizer in the garden and spray insectifuge to fully cover both potted plants in the garden.

Result

Mean q_score	Full successes
0.000	0 / 10

For reference, 2025 leaderboard on this task: Comet 0.15, Robot Learning Collective 0.20 (from the eval kit).

Instance	Instance ID	q_score	Success	Steps	Video	Frames
0	301	0.000	✗	9720	video	frames
1	302	0.000	✗	9720	video	frames
2	303	0.000	✗	9720	video	frames
3	304	0.000	✗	9720	video	frames
4	305	0.000	✗	9720	video	frames
5	306	0.000	✗	9720	video	frames
6	307	0.000	✗	9720	video	frames
7	308	0.000	✗	9720	video	frames
8	309	0.000	✗	9720	video	frames
9	310	0.000	✗	9720	video	frames

Every rollout ran to the timeout (9,720 steps = 1.5× mean human demo length, 324 s simulated).

What the videos show

• The policy navigates along the garden path toward the pesticide atomizer (sometimes after wandering for a while).
• It then fails to grasp it: the left gripper hovers over the atomizer for the rest of the episode, often knocking it over.
• It never carries the atomizer to the potted plants or sprays, so no BDDL goal predicate is satisfied and partial credit is 0.

Setup

• Evaluator: python -m omnigibson.eval.eval --task-name spraying_for_bugs --mode public_test --instance-indices <i> --num-envs 1 --num-rollouts 1 --host 127.0.0.1 --port 8000 --output-dir <out> --write-video (default wrapper, r1pro.yaml)
• Policy server: OpenPI serve_b1k.py --robot b1k/R1Pro_comet --task b1k/spraying_for_bugs --policy.config pi05_b1k_spraying_for_bugs_comet_lora --control_mode receding_horizon --action_horizon 16
• Hardware: 1× NVIDIA L40S 46 GB, ~33 min wall-clock per rollout.
• Script: config/run_eval.sh (from the model repo's eval_kit).

Fixes needed for v3.9.3 (applied here)

1. Robot name: v3.9.3 names the robot robot_r1 (was robot), so observation keys are robot_r1::proprio and robot_r1::robot_r1:<camera>::rgb. The OpenPI robot config was updated; see config/robot_name_fix.patch. The proprio index layout is unchanged. The first run without this fix crashed with KeyError: 'robot::proprio' (logs in logs/failed_run1/, no results).
2. og_smoke_test.py: also set gm.ENABLE_TRANSITION_RULES = False (a scene with transition rules asserts object states are on).
3. summarize_results.py: q_score is now a dict ({"final": …}); results/summary.json was produced directly instead.
4. Headless container: Vulkan needed the EGL ICD (VK_ICD_FILENAMES → libEGL_nvidia.so.0).

Files

• results/json/: official per-rollout metrics (q_score, agent distance, time)
• results/videos/: rollout MP4s (head + wrist cameras)
• results/summary.json: mean q_score and per-rollout table
• previews/: 8 evenly spaced frames per rollout
• logs/: evaluator, policy-server and run logs
• config/: evaluator robot config, patched OpenPI robot config, eval scripts

Note: this is a single-task self-evaluation, not a full 100-task challenge submission.