Intended Hub repository identity: ASethi04/pi05-BimanualYAM-yam-ultrawide-teleop-ee20-measured-fk-12k. This directory is a verified local staging artifact; the staging tools perform no Hub write. Artifact profile: with-training-config. This profile includes a sanitized train_config.json for training reproducibility.

This policy was fine-tuned for 12,000 optimizer steps with global batch 64 on 75 YAM teleoperation episodes. The exact task is pick up oranges and place them in the bowl. Inputs are the current left and right wrist images plus the current identity-pose/current-gripper EE20 state. The absent Pi0.5 base camera slot remains masked.

Action contract

The output shape is (24, 20). Each arm contributes xyz (3) + first two rotation-matrix rows (6) + absolute normalized gripper (1), left then right. Every future row is independently query anchored:

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

The training labels use measured YAM joints and authoritative I2RT FK at both the query and future rows. At deployment, PR #15 reconstructs T_target = T_query @ DeltaT, runs I2RT IK independently per arm, and fails closed on its unchanged residual, joint-limit, and frozen-query seed gates. Grippers are model-space [0, 1] scalars (0 closed, 1 open); the deployment adapter alone applies the 85 mm / 96 mm driver mapping. No legacy tool-axis transform is used.

Immutable lineage

Source: brandonyang/yam-ultrawide-teleop@e0fae691e2eee74430dd463adf8e17180bc735e9 (108,726 frames).
Derived measured-FK view: 106,926 valid H24 queries; payload manifest be8b88a612c6677e142957e3b8c223d0ee0d08434e65abfda51ab33d9931d6f9.
Base: lerobot/pi05_base@b211f3d44c36b6acfcf7ae94a64e8e96f75a64ba.
Tokenizer: google/paligemma-3b-pt-224@35e4f46485b4d07967e7e9935bc3786aad50687c. The inference host must accept its terms and authenticate to Hugging Face.
I2RT: 7ed46f4e4e316133a0c39aa6cf34a73d2718e850, YAM + LINEAR_4310, authoritative grasp_site.
Training code: 4e439e9e62c9a65a58d22bafe5efcd4cba12f1cf.
Deployment code: 39af035ae1b9007c9c6d9c290d21fae8cb58b802.

Machine-readable hashes and the exact sanitization record are in artifact_provenance.json; verify every file with SHA256SUMS before use.

Limitations and safety

No held-out split was created. Any offline replay is an in-domain training-set replay and is not evidence of generalization, closed-loop success, or hardware safety. Offline model error and I2RT adapter acceptance are separate diagnostics. A passing IK row does not replace collision checking, calibration, workspace limits, an emergency stop, or a low-speed no-motion/action probe.

Use the clean deployment code at 39af035ae1b9007c9c6d9c290d21fae8cb58b802, the exact prompt above, H24 inference, and the hardware team's approved execution horizon. Pin an immutable Hub commit after publication; never deploy an unpinned mutable revision.