Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern_VLA-JEPA-CurrentState-QwenTrain-WMOn_bs16_step8000

VLA-JEPA Current-State Qwen-Train WM-On - Task 000004

LeRobot 0.6.1 inference checkpoint at step 8,000, trained on all 35 episodes of Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern revision 05286a17a145234ed80870702f4d9757f00194c3.

• —Batch size: 16
• —State/action: 22 dimensions
• —Cameras: head, left wrist, right wrist
• —Action chunk / executed actions: 7 / 7
• —Qwen trainable (freeze_qwen=false)
• —World model enabled during training
• —V-JEPA2 visual encoder frozen; action-conditioned video predictor trainable
• —Legacy LeRobot 0.6.1 world-model context (causal_world_model_context=false equivalent)
• —Corrected current-state input path (state[:, 0, :])
• —Normalization: state MEANSTD, action MINMAX, visual IDENTITY

This repository contains inference files only. Optimizer, scheduler, and RNG state remain in the local training checkpoint. The resumed 8k save omitted serialized processor state; the processor JSON and normalization tensors here are restored byte-for-byte from the 5k checkpoint of the same run and dataset. Model weights and model/train configs come from 8k.