albertwilcox/bimesa-57-lerobot

BiMESA-57 (LeRobot)

BiMESA-57 is the bimanual training set of MESA: 57 tasks and 6,833 simulated demonstrations on two ReverseMountedYam arms (robot0 = left, robot1 = right), in LeRobot v2.1 format.

Block	Skill	Tasks	Demos
A	Spatial pick-and-place	27	4,690
B	Cross-arm handoff	17	843
C	Single-arm articulated	13	1,300
Total		57	6,833

Block A aligned combos and all of Block C are generated with MESA-Gen; Block A cross combos and all Block B handoffs are teleoperated. Held-out combos, reversed handoff directions, and fixture sides are reserved for the BiMESA evaluation suites (bimesa-id, bimesa-spatial, bimesa-instance, bimesa-object, bimesa-composite); see the documentation.

Format

1,116,641 frames at 20 fps. Features:

Key	Shape	Description
observation.images.egocentric	224x224x3 (video)	egocentric camera
observation.images.robot0_eye_in_hand	224x224x3 (video)	left wrist camera
observation.images.robot1_eye_in_hand	224x224x3 (video)	right wrist camera
observation.state	14	per arm: 6 joint positions + gripper width, robot0 then robot1
action	14	absolute joint-position targets, same layout as the state

The language instruction for each episode is its LeRobot task (meta/tasks.jsonl).

Usage

Policies trained on this dataset (π0, π0-FAST, π0.5) are released in the MESA collection; the corresponding training configs are in pairlab/openpi-mesa (pi0_bimesa, pi0_fast_bimesa, pi05_bimesa).

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}
}