MolmoAct2-BimanualYAM-freshbase-rotcontract-umi75-teleop25-ee20
Fresh-base molmoact2 EE20/H24 research checkpoint, trained for 12,000 optimizer steps with seed 1000.
- UMI share: 75%
- teleop share: 25%
- UMI dataset:
brandonyang/dual-lidar-umi-independent@a95b079b2b3dc73a912ecd12967f22f825d04fa8
- teleop dataset:
brandonyang/yam-ultrawide-teleop@e0fae691e2eee74430dd463adf8e17180bc735e9
- task:
pick up oranges and place them in the bowl
- rotation contraction: left=0.6034407436097629, right=0.5799117511664429 (UMI orientation targets only)
- action per arm:
xyz + first two rotation-matrix rows + absolute future normalized gripper
- transform:
inverse(T_t) @ T_(t+k) independently for k=1..24
- arm order: left, then right
- no terminal padding; execute exactly H24
The SO(3) contraction is part of the learned UMI target, not a frame conversion.
Execute the predicted R6D rotations directly through the normal EE-to-IK path.
Do not divide predicted rotation vectors by the scales or otherwise decontract them before IK.
This is a research checkpoint. Hardware use still requires the deployment-side IK, joint/collision limits, I2RT success checks, and the normal physical safety procedure. Pin the immutable Hub revision rather than main.
