ASethi04/MolmoAct2-BimanualYAM-dual-lidar-umi-currentrel-r6d-onset-v3-12k

MolmoAct2 Bimanual YAM: dual-lidar UMI onset-v3 (12k)

MolmoAct2 policy fine-tuned for bimanual UMI grippers mounted on YAM, using current-relative dual end-effector trajectories.

• Training dataset: brandonyang/dual-lidar-umi-currentrel-r6d-onset-v3
• Training steps: 12,000
• Seed: 1,000
• Action chunk: 24 steps, 20 action dimensions
• Visual observations: two 600x800 UMI camera streams
• Action mode: continuous flow matching with 8 flow timesteps

The repository contains the LeRobot policy config, safetensors weights, training config, and matching policy preprocessor/postprocessor artifacts from checkpoint 012000.

Loading

Use LeRobot with MolmoAct2 support and load this repository as the policy pretrained path.