OpenVLA-OFT
Optimized fine-tuning recipe for OpenVLA: parallel action-chunk decoding, continuous L1 regression head and FiLM language conditioning give 26x faster inference and near-saturated LIBERO scores.
- 97%
- 34K
- 25 Hz
- 62K
Benchmarks
Numbers as reported by the authors on the named evaluation suite.
| Evaluation | Suite | Score |
|---|---|---|
LIBERO average | LIBERO (4 suites) | 97.1% |
ALOHA real-world tasks | OFT eval | 78% |
Inference throughput | A100 | 26 |
Training
OpenVLA + parallel decoding & continuous actions
7B
8x A100 80GB
1.6K accelerator-hours
1.4K h
62K
L1 action-chunk regression instead of token classification; FiLM injects language into the vision tower.
Chunk size and proprioception inputs matter more than backbone size for bimanual tasks.
Linked datasets, robots & papers
Resolving related models, datasets and papers…