OpenVLA 7B
Fully open 7B vision-language-action model trained on 970k real robot episodes. Beats RT-2-X by 16.5% absolute on 29 tasks with 7x fewer parameters, and fine-tunes on one A100 with LoRA.
- 71%
- 96K
- 5 Hz
- 970K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
BridgeData V2 (WidowX) | OpenVLA eval | 71% |
Google robot generalization | OpenVLA eval | 76% |
vs RT-2-X | 29 tasks, 2 embodiments | 16.5 |
LIBERO-Long Base checkpoint, no fine-tune | LIBERO | 53.7% |
Training
Prismatic VLM (Llama-2 7B + DINOv2/SigLIP)
7B
64x A100 80GB
22K accelerator-hours
19K h
970K
27 epochs of next-action-token prediction, fused DINOv2 + SigLIP features, action bins from per-dim quantiles.
LoRA rank 32 fine-tunes on a single A100 in ~10 h; 4-bit quantized inference loses <1% success.
Linked datasets, robots & papers
Resolving related models, datasets and papers…