MolmoAct2-BimanualYAM

Ai2's open-source bimanual teleoperation dataset for MolmoAct2, totaling more than 720 hours.

MolmoAct2-BimanualYAM is a bimanual-manipulation dataset the Allen Institute for AI (Ai2) open-sourced in May 2026 alongside its MolmoAct2 model, collected and curated with support from Cortex AI. The data consists of teleoperated demonstrations using two I2RT YAM robot arms (6 joints plus 1 gripper per arm, giving a 14-dimensional joint-position action space), totaling more than 720 hours; the merged version on Hugging Face contains 32,246 trajectories and about 76 million frames (at 30 frames per second), mostly tabletop bimanual tasks such as folding a towel, scanning items, and charging a phone, with each trajectory carrying an annotated language instruction. Ai2 describes it as the largest open bimanual dataset to date, more than 30 times the amount of robot data used for the original MolmoAct. The data uses the LeRobot v3.0 format under the Apache 2.0 license, and can be used directly to pretrain or fine-tune a bimanual VLA (vision-language-action model).

ExampleMolmoAct2 used this dataset as one of its main training sources, mixed with other robot datasets; Ai2 says the resulting model can perform bimanual manipulation without needing separate fine-tuning for each task.

Also called

MolmoAct2-BimanualYAM Dataset, MolmoAct2-Bimanual YAM

Related

MolmoAct · I2RT YAM Arm · Bimanual Manipulation · Teleoperation · LeRobotDataset · Allen Institute for AI

Sources

Ai2 博客：MolmoAct 2 (Chinese)
allenai/MolmoAct2-BimanualYAM-Dataset（Hugging Face）
MolmoAct2: Action Reasoning Models for Real-world Deployment (arXiv 2605.02881)

As of

2026-06