Pi0.5 BiYAM EE20: rotation-contracted UMI 75% / teleop 25%
Research checkpoint for hardware comparison. The policy starts from the successful 12k teleop-to-EE checkpoint and is adapted for 4,000 steps with a 75% UMI / 25% teleop mixture. UMI rotation magnitudes are contracted toward the teleop distribution; translations, grippers, and all teleop targets are unchanged.
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.
This is an experimental comparison model, not the primary teleop baseline. Run normal collision, workspace, joint-limit, IK-success, and operator-stop gates.
Training sources:
- UMI:
brandonyang/dual-lidar-combined-filtered-long-gripper@a29ae6a5531584fb950c7bb3bb5895f18421b108 (older 182-episode set)
- Teleop:
brandonyang/yam-ultrawide-teleop@e0fae691e2eee74430dd463adf8e17180bc735e9
- UMI rotation contraction: left
0.603, right 0.559
- Adaptation: 4,000 steps, global batch 64, learning rate
5e-6
- PaliGemma/tokenizer revision:
google/paligemma-3b-pt-224@35e4f46485b4d07967e7e9935bc3786aad50687c
Offline held-out replay found finite H24 outputs and valid R6D rows for both arms. That is an engineering sanity check, not evidence of task success.
