This is the 20-dimensional end-effector member of the four-model UMI/YAM comparison. It was fine-tuned for 12,000 optimizer steps with global batch 64 on 182 bimanual demonstrations.

Physical motion with this checkpoint has not yet been validated. Start with the documented no-motion action_probe, review its output, and execute one predicted row per policy query.

Training lineage

Base: lerobot/pi05_base at b211f3d44c36b6acfcf7ae94a64e8e96f75a64ba.
Cartesian data: brandonyang/dual-lidar-combined-filtered-long-gripper at a29ae6a5531584fb950c7bb3bb5895f18421b108.
Paired normalized grippers: brandonyang/dual-lidar-combined-filtered-joint-positions-long-gripper at 387d696eb36de411a11c8b2612a6f19822ca3a53.
Data: 182 episodes, two UMI wrist cameras (umi1 left, umi2 right), 800×600 at 30 fps.
Task: pick up oranges and place them in the bowl.
Prediction horizon: 24 rows (0.8 seconds); recommended initial execution horizon: 1 row.
Pi0.5 uses its continuous flow-matching action head. Componentwise relative-action conversion is disabled because EE20 is already SE(3)-relative.

Action contract

The model predicts a tensor of shape (24, 20). Each arm contributes:

relative xyz (3) + first two rotation-matrix rows (6) + absolute normalized gripper (1)

For query time t, every future row is independently anchored to the measured query pose:

DeltaT[k] = inverse(T[t]) @ T[t+k], for k = 1..24.

At runtime, the deployment adapter measures both arms, obtains each UMI-compatible grasp pose through I2RT FK, reconstructs T_target = T_query @ DeltaT, and runs I2RT IK independently for the left and right arms. It then emits one validated 14-dimensional joint command. This model therefore uses online IK; the joint14 model does not.

Kinematics use AfterQuery I2RT 7ed46f4e4e316133a0c39aa6cf34a73d2718e850, including the 220 mm flange-to-grasp offset and UMI-compatible X-up, Y-right, Z-forward frame. No legacy axis transform is applied.

Deployment

Use the deployment code and safety procedure in AfterQuery-Research/lerobot PR #15. Pin the exact Hub commit when serving:

bash

uv run lerobot-policy-server \
  --policy.pretrained_name_or_path=ASethi04/pi05-BimanualYAM-long-gripper182-ee20 \
  --policy.revision=<immutable-hub-commit> \
  --policy.device=cuda \
  --host=0.0.0.0 \
  --port=8081

Use --inference.bi_yam_action_mode=ee, the exact task text above, and --inference.execution_horizon=1. Before enabling motion, run the no-motion action_probe from the PR runbook and verify current per-arm motor calibration.

The tokenizer is pinned to google/paligemma-3b-pt-224 at 35e4f46485b4d07967e7e9935bc3786aad50687c. The inference host must accept that repository's terms and authenticate to Hugging Face; no Harvard filesystem path is required.