Is Forward Prediction Enough? Physical State Grounding for JEPA World Models
Paper • 2608.06799 • Published • 1
PSG-JEPA — LIBERO-Goal checkpoints
Checkpoints for the LIBERO-Goal policy-learning result in "Is Forward Prediction Enough? Physical State Grounding for JEPA World Models" (arXiv:2608.06799).
Code, evaluation logs and the download helper are in
Haodong-Yan/PSG-JEPA under libero/.
| File | What it is |
|---|---|
psgjepa/seed3072/encoder_epoch40.ckpt |
PSG-JEPA encoder, 40 pretraining epochs |
psgjepa/seed3072/oft_head_epoch30.ckpt |
OFT action head, jointly fine-tuned with the encoder |
Training seed 3072. Closed-loop evaluation of the head over 10 LIBERO-Goal tasks x 50 rollouts (eval seed 4242, max_steps 600, action horizon 8) gives 90.8% mean success.
The baselines are reproduced by running the pipeline in the code repository; their evaluation
logs ship there under libero/results/.
Download
python libero/weights/download_weights.py # verifies both files against manifest.json
libero/weights/manifest.json records the sha256 and size of each file.