CoolFace
Modelpublic

Dongkkka/vla_jepa_dashboard_0904_5k_8bs

sourceHugging Faceupdated 16d agoView on Hugging Face
0likes22downloads
Model Card

VLA-JEPA dashboard_0904: 5,000 steps, batch 8

5,000-step checkpoint selected by the lowest held-out active-range-normalized MAE among the 2k, 5k, 10k and 20k saves of a 20,000-step run. Held-out episodes 24-27: 699 frames; arm-joint MAE 0.0582105; active-range-normalized MAE 0.1086933. These are dataset open-loop metrics, not robot task-success rates. The saved training config retains the total 20k run target.

[image] Fine-tuned from lerobot/VLA-JEPA-Pretrain revision e946c3e5b538d760f4b4ff239d1b1c12090c041d.

  • —Dataset: Dongkkka/cyclo_dashboard_0904_test_v30.
  • —Training episodes: 0-23; held-out episodes: 24-27.
  • —State/action dimensions: 22 each. Action order: left arm 7, left gripper, right arm 7, right gripper, head 2, lift 1, base velocities 3.
  • —Cameras: observation.images.rgb.cam_left_head, observation.images.rgb.cam_left_wrist, observation.images.rgb.cam_right_wrist.
  • —Qwen input resize: 224 x 224 per camera.
  • —Action chunk / execution length: 7 / 7; flow inference steps: 4.
  • —State normalization: MEANSTD; action normalization: MINMAX.
  • —Qwen backbone unfrozen; world-model training enabled.
  • —Gripper snap and binarization disabled.
  • —Cross-embodiment projection initialization was performed before training; this repository contains the trained 22D weights.

Model weights, config, saved pre/postprocessors and normalization state files are at the repository root. Apply both saved processors when using the policy; raw policy output can be normalized. The training config is provided for provenance, with the local dataset and output paths removed/replaced. Optimizer/RNG state is not included; this is a policy export, not a full training-resume bundle.

Training used Cyclo's installed LeRobot 0.6.1 environment with compatibility changes, Transformers 5.5.4 and PyTorch 2.11.0+cu128. Install compatible VLA-JEPA dependencies when loading. Qwen/V-JEPA processor assets may be fetched by the native loader. Robot rollout performance is not guaranteed by this upload.