MolmoAct2 BiYAM EE20: canonical UMI 75% / teleop 25%
Research checkpoint for hardware comparison. This MolmoAct2 policy starts from the successful 12k measured-teleop-to-EE checkpoint and trains to step 12,000 with the raw canonical 75% UMI / 25% teleop mixture. No rotation contraction or other label transport is applied.
The policy predicts an H24 chunk of EE20 actions. Each arm contributes xyz + R6D (first two rotation-matrix rows) + normalized absolute future gripper, ordered left then right. Every future pose is query-relative: DeltaT_k = inverse(T_t) @ T_(t+k). Use the exact task prompt pick up oranges and place them in the bowl and execute all 24 actions. Robot-side deployment converts each relative EE target through I2RT IK.
Training sources:
- Canonical UMI:
brandonyang/dual-lidar-umi-independent@a95b079b2b3dc73a912ecd12967f22f825d04fa8
- Measured-FK teleop:
brandonyang/yam-ultrawide-teleop@e0fae691e2eee74430dd463adf8e17180bc735e9
- Exact queries: 269,228 UMI + 89,743 teleop
- Global batch 64; 12,000 adaptation steps
- FAST tokenizer:
allenai/MolmoAct2-FAST-Tokenizer@d45593b4c863d0bc1ca064f8b352fa16b75c38e8
This is an experimental comparison model. Use normal collision, workspace, joint-limit, IK-success, and operator-stop gates.
