Skip to content

Reproduction Issue: Extremely low Q-score (~0.03 / 3%) on BEHAVIOR-1K 50 tasks with pi05-b1kpt50-cs32 despite correct setup; latest main branch causes multiple import/Hydra errors #24

Description

@riyuu2197

Problem Description

I followed the official README of openpi-comet to reproduce the 50-task BEHAVIOR-1K results using the provided pi05-b1kpt50-cs32 checkpoint (the one specifically trained for the 50 tasks).

Expected results (according to leaderboard + paper):

Validation success rate ≈ 0.1440 / Q-score ≈ 0.1830 (submission value)
Post-competition optimized Q-score ≈ 0.345

Actual results:

On BEHAVIOR-1K v3.7.0 and v3.7.2 (stable releases): Q-score only ~0.03 (3%) across 50 tasks.
On latest main branch (as of 2026-04-14): evaluation fails with a series of import / NameError / Hydra errors (detailed below).
Reproduction Steps (strictly followed README)

Cloned openpi-comet (latest main) and BEHAVIOR-1K.
Copied all custom files:
cp -r openpi-comet/src/behavior/learning/* BEHAVIOR-1K/OmniGibson/omnigibson/learning/
Started policy server (confirmed checkpoint + norm_stats loaded successfully):

CUDA_VISIBLE_DEVICES=3 python scripts/serve_b1k.py
--task_name=turning_on_radio
--control_mode=receeding_horizon
--max_len=32
policy:checkpoint
--policy.config=pi05_b1k-base
--policy.dir=../pi05-b1kpt50-cs32

Ran evaluation with correct flags:

python OmniGibson/omnigibson/learning/eval.py
policy=websocket
task.name=turning_on_radio
log_path=./logs/turning_on_radio_test
env_wrapper.target=omnigibson.learning.wrappers.RGBWrapper

Errors on latest main branch

Even after commenting out all non-essential imports (lerobot, gello, BaseRobot, register_omegaconf_resolvers, etc.) as suggested in the debugging process, we still hit:

ModuleNotFoundError: No module named 'lerobot.datasets...'
NameError: name 'HF_LEROBOT_HOME' is not defined
ModuleNotFoundError: No module named 'gello.robots.sim_robot.og_teleop_cfg'
ImportError: cannot import name 'BaseRobot' from 'omnigibson.robots'
NameError: name 'BaseRobot' is not defined (in type hints)
hydra.errors.ConfigCompositionException: You must specify 'policy'... (even when policy=websocket is explicitly passed)
NameError: name 'register_omegaconf_resolvers' is not defined

On stable v3.7.x the evaluation runs but success rate stays near zero.

Environment

openpi-comet: latest main
BEHAVIOR-1K: tested v3.7.0, v3.7.2, and latest main
Python: 3.10 (behavior conda env)
Checkpoint: official pi05-b1kpt50-cs32
GPU: RTX 5090 (serve logs show successful JAX/Orbax checkpoint restore)

Request

Could the team please clarify:

Which exact BEHAVIOR-1K commit / tag is officially supported with the current openpi-comet custom files (the 2026-01-06 version)?
Is there a recommended stable commit hash for BEHAVIOR-1K that matches the pi05-b1kpt50-cs32 weights?
Any additional flags or post-processing steps needed to reach the reported Q-score of ~0.345?
Happy to provide full logs, serve output, or test any suggested commit. Thank you!

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions