Modelpublic

Dongkkka/Task_000004_Peanut_Pick_Place_lerobot_Intern_VLA-JEPA_bs16_step12000

sourceHugging Faceupdated 9d agoView on Hugging Face

0likes19downloads

Model Card

VLA-JEPA — Task 000004 Peanut Pick & Place

—LeRobot version: 0.6.1
—Batch size: 16
—Published checkpoint: step 12000 (selected best validation checkpoint at upload time)
—Validation loss at step 12000: 0.1452
—Training is ongoing; this is not the final training checkpoint.
—Dataset revision: 05286a17a145234ed80870702f4d9757f00194c3
—Cameras: camlefthead, camleftwrist, camrightwrist
—State/action dimensions: 22
—Training/evaluation image resize: 224 x 224 (antialias=True); this dataset-side transform is outside the saved processor pipeline.
—State normalization: MEANSTD; action normalization: MINMAX.

This repository contains the unchanged step-12000 pretrained model, train configuration, and saved pre/postprocessor configurations and normalization statistics. Optimizer state and other checkpoints are excluded.

For inference, load both saved policypreprocessor.json and policypostprocessor.json with their referenced safetensors statistics. Normalize observations before the policy and unnormalize its actions afterwards. The server-local JEPA factory override currently rebuilds processors without statistics when dataset_stats is omitted: uploading these files alone does not fix that loader. Load the saved processor pipelines directly.

Recorded-data open-loop evaluation (CPU, 50 uniformly sampled frames per episode, first action per predicted chunk, step-12000 weights):

—Held-out left episode 8 MAE: 0.0226846971
—Held-out right episode 12 MAE: 0.0275056921
—Mean of the two held-out episode MAEs: 0.0250951946
—Static episode 4 MAE: 0.0032854648 (included in training; not held-out) These raw action-unit metrics cover a small sample and do not establish closed-loop robot performance. No smoothing or weight modification was applied.