kamusarj/VLA-JEPA_finetune

VLA-JEPA UR10e fine-tune — 17,500 steps

This repository contains the final model-only checkpoint from the VLA-JEPA UR10e joint action/world-model fine-tuning run. Training used the canonical LeRobot v2.1 artifact ur10e-cup-lerobot-v2.1-v1 and stopped at exactly 17,500 optimizer steps.

Training summary

• Architecture: VLA-JEPA with Qwen3-VL-2B, V-JEPA2 target encoder/predictor, and a DiT-B action head
• Objective: action loss + 0.1 * world-model loss
• Training episodes: 0–72
• Holdout episodes: 73–80
• Video horizon: 8 frames
• Action horizon: 7
• Seed: 42
• World-model gradient check: passed for Qwen-VL interface, V-JEPA predictor, and action head; the V-JEPA target encoder remained frozen

The run reached 10,000 steps first and then continued to 17,500 from the model-only checkpoint. The optimizer was reinitialized for the continuation, and the learning-rate schedule was reconstructed at step 10,000. This is not an optimizer-resumable checkpoint.

Open-loop evaluation

The final checkpoint was evaluated offline on all eight holdout episodes using the GR00T-style chunked open-loop protocol (execution horizon 7, eight denoising steps).

Metric	Model	Persistence baseline
Global unnormalized MSE	0.001192	0.001966
Global unnormalized MAE	0.003733	0.005706
Moving-joint MAE (rad)	0.003553	0.005338
Gripper accuracy	99.186%	98.677%

These results measure offline next-action prediction, not closed-loop robot task success or recovery behavior.

Files

• pytorch_model.pt: model weights at step 17,500
• config.yaml: effective training configuration
• artifacts/ur10e_step17500/: training history, metadata, evaluation metrics, plots, and raw per-episode rollout arrays
• open_loop_report.md: human-readable evaluation report

Checkpoint integrity:

size:   6,826,992,518 bytes
sha256: 313e4a8e764a7482aa90b54364b9c898119185212687a3b880e5cda60cfa6094

Load the checkpoint using the VLA-JEPA codebase at https://github.com/kamusarj/VLA-JEPA. Exact training and evaluation provenance is preserved in the JSON files under artifacts/ur10e_step17500/.