lerobot / hy_vla_robotwin

Hy-Embodied-0.5-VLA — LeRobot conversion

Byte-identical tensor conversion of tencent/Hy-Embodied-0.5-VLA-RoboTwin at revision bd7bba6f5934ad62293a2a34f74760c6a3ef2ff8 .

• Author runtime revision: 8ba4c8cbdf42a4bcf0a19be4bd2841405dfe15e9
• Model token horizon: 40
• Physical action horizon: 20
• Actions executed per sample: 7
• Image history: 6 frames at interval 5
• Action representation: relative_absolute
• Euler inference steps: 10
• Padded state/action width: 32/32
• Tensor mapping: identity (1628 tensors)

The runtime consumes an already-selected LeRobot task string and never samples task augmentation text. norm_stats.pkl is retained for provenance; normalization is loaded at runtime from safe serialized processor state.

Validation

• Strict load, save_pretrained , and strict reload passed with all 1,628 tensors mapped and no missing, unexpected, duplicate, or shape-mismatched tensors.
• Full author-vs-LeRobot forward parity passed for batch sizes 1 and 2. Images, masks, tokens, state/targets, training velocity/loss, every velocity from the 10-step Euler sampler, the complete 11-state trajectory, and final sampled actions all had max, p95, and p99 error 0.0 .
• Final relative, absolute, and blended 16D RoboTwin actions had scalar and rotation-matrix max/p95 error 0.0 .
• On the pinned first-six-task RoboTwin quick protocol, the author wrapper scored 60/60. LeRobot scored 53/54 completed episodes; conservatively counting one native-planner hang as a failure gives 53/55 (96.36%) versus the matched author subset's 55/55, a 3.64-percentage-point delta.

Limitation

This policy supports the released 16D dual-arm pose/gripper embodiment only. A 12D mobile-manipulator action contains base semantics that the released 20D dual-arm representation cannot express and is rejected rather than silently truncated or reinterpreted.