metadata
license: gemma
datasets:
- allenai/molmobot-data
language:
- en
base_model:
- google/paligemma-3b-mix-224
pipeline_tag: robotics
tags:
- robotics
- manipulation
MolmoBot-Pi0-DROID
[Paper] [Project Website] [Code] [Data]
MolmoBot-Pi0-DROID is the MolmoBot-Pi0 VLA trained on simulation data on the DROID platform, without any real robot data. See here for usage instructions.
Quickstart
Note that the official policy configuration used in our evals is fully detailed here.
import numpy as np
from huggingface_hub import snapshot_download
from molmobot_pi0.eval.policies.pi import PiJointPosPolicy
ckpt_dir = snapshot_download("allenai/MolmoBot-Pi0-DROID")
policy = PiJointPosPolicy(checkpoint_dir=ckpt_dir)
policy.prepare_model()
obs = {
"task": "put the mug in the bowl",
"qpos": {
"arm": np.zeros(7),
"gripper": np.zeros(2),
},
"exo_camera_1": np.zeros((360, 640, 3), dtype=np.uint8),
"wrist_camera": np.zeros((360, 640, 3), dtype=np.uint8),
}
action = policy.get_action(obs)
print(action)
BibTeX
@misc{deshpande2026molmobot,
title={MolmoB0T: Large-Scale Simulation Enables Zero-Shot Manipulation},
author={Abhay Deshpande and Maya Guru and Rose Hendrix and Snehal Jauhri and Ainaz Eftekhar and Rohun Tripathi and Max Argus and Jordi Salvador and Haoquan Fang and Matthew Wallingford and Wilbert Pumacay and Yejin Kim and Quinn Pfeifer and Ying-Chun Lee and Piper Wolters and Omar Rayyan and Mingtong Zhang and Jiafei Duan and Karen Farley and Winson Han and Eli Vanderbilt and Dieter Fox and Ali Farhadi and Georgia Chalvatzaki and Dhruv Shah and Ranjay Krishna},
year={2026},
eprint={2603.16861},
archivePrefix={arXiv},
primaryClass={cs.RO},
url={https://arxiv.org/abs/2603.16861},
}