albertwilcox/bimesa-pi0

BiMESA π0

π0 (flow-matching action expert on a PaliGemma-3B backbone, pretrained by Physical Intelligence) finetuned on BiMESA-57 (57 tasks, 6,833 demonstrations; LeRobot v2.1) for the bimanual (two ReverseMountedYam arms) setting of MESA. This is the checkpoint behind the π0 BiMESA results in the MESA paper.

This is an openpi (JAX/orbax) checkpoint; use it with our fork, pairlab/openpi-mesa, config pi0_bimesa.

Model details

Base model	π0 (gs://openpi-assets/checkpoints/pi0_base)
openpi-mesa config	pi0_bimesa
Training data	albertwilcox/bimesa-57-lerobot
Training	50k steps, batch size 128, default openpi optimizer; final checkpoint (step 49999)
Cameras	egocentric camera (egocentric) and both wrist cameras (robot0_eye_in_hand, robot1_eye_in_hand), 224x224 RGB
Proprioception	14-D: per arm, 6 joint positions + gripper width, robot0 then robot1
Actions	14-D absolute joint-position targets (per arm: 6 joints + gripper, robot0 then robot1). The model predicts joint positions relative to the current state (grippers absolute); the policy's output transforms convert them back to absolute targets.
Action chunk	20 steps at 20 Hz; the MESA evaluation server executes the first 5 actions before replanning
Language	task instruction
Normalization stats	assets/bimesa/norm_stats.json

Usage

Install openpi-mesa following its README, then download and serve the checkpoint:

uv run huggingface-cli download albertwilcox/bimesa-pi0 --local-dir checkpoints/bimesa-pi0
uv run scripts/serve_policy.py --port 8001 policy:checkpoint \
  --policy.config=pi0_bimesa \
  --policy.dir=checkpoints/bimesa-pi0

From the MESA repository, run the evaluation server against the same port:

uv run scripts/eval_server_parallel.py \
  --port 8001 \
  --eval-set-name bimesa-id \
  --num-rollouts-per-task 50 \
  --controller-type joint_pos \
  --robots ReverseMountedYam ReverseMountedYam \
  --camera-names egocentric robot0_eye_in_hand robot1_eye_in_hand \
  --state-keys robot0_joint_pos robot0_gripper_jaw_width robot1_joint_pos robot1_gripper_jaw_width

Evaluation suites: bimesa-id, bimesa-spatial, bimesa-instance, bimesa-composite, bimesa-object. See the MESA documentation for details.

Results

Success rates (%) from the MESA paper (Table 2), 50 rollouts per task:

Suite	Success rate
BiMESA-ID	45.9
BiMESA-Spatial	49.2
BiMESA-Instance	23.9
BiMESA-Composite	18.4
BiMESA-Category	10.6
Average	29.6

Notes

• The checkpoint contains the model parameters (params/), the model config (config/), and normalization stats (assets/); optimizer state is not included.

License

Released under the Apache 2.0 license. The model is finetuned from openpi weights that build on PaliGemma; use of the weights may also be subject to the Gemma Terms of Use.

Citation

@inproceedings{
wilcox2026mesa,
title={{MESA}: An Evaluation Framework for Compositional, Semantic, and Spatial Generalization in Robotics},
author={Wilcox, Albert and Chang, Frank and Nguyen, Nhi and Chakraborty, Aishani and Collins, Jeremy A. and Saxena, Vaibhav and Joffe, Benjamin and Karamcheti, Siddharth and Garg, Animesh},
booktitle={10th Annual Conference on Robot Learning},
year={2026},
url={https://openreview.net/forum?id=Br2rXixvyN}
}