EgoGym

A lightweight benchmark suite for evaluating CAP policies and π_0.5 zero-shot.

import gymnasium as gym
import egogym

env = gym.make(
    "Egogym-Pick-v0",           # Options: "Egogym-Pick-v0", "Egogym-Open-v0", "Egogym-Close-v0"
    robot="cap",                # Options: "cap", "droid"
    action_space="delta",       # Options: "delta", "absolute"
    num_objs=5,                 # Options: 1-5
)

Run a baseline

# CAP
python run.py --policy baselines/cap_policy.py --task pick --num_objs 5 \
  policy.model_path=checkpoints/cap-pick.pt

# π_0.5 (after running openpi server)
python run.py --policy baselines/pi_policy.py --task pick --num_objs 5 \
  --action_space absolute policy.host=0.0.0.0 policy.port=8000

Citation

If you use EgoGym in your work, please cite:

@article{cui2026contact,
  title={Contact-Anchored Policies: Contact Conditioning Creates Strong Robot Utility Models},
  author={Cui, Zichen Jeff and Rayyan, Omar and Etukuru, Haritheja and Tan, Bowen and Andrianarivo, Zavier and Teng, Zicheng and Zhou, Yihang and Mehta, Krish and Wojno, Nicholas and Wu, Kevin Yuanbo and others},
  journal={arXiv preprint arXiv:2602.09017},
  year={2026}
}