Paper
OpenVLA: An Open-Source Vision-Language-Action Model
External resource. Not made or verified by Public Protocol.
A 7-billion-parameter open vision-language-action model trained on 970,000 real robot demonstrations (June 2024). The authors report 16.5 percent higher absolute task success than the 55-billion-parameter RT-2-X across 29 tasks, fine-tuning on consumer GPUs with low-rank adaptation, and no loss of success from quantised serving.
https://arxiv.org/abs/2406.09246
- Maintainer
- Kim, Pertsch and co-authors
- Licence, as stated at its source
- CC BY 4.0 (paper); model licence not stated on the page opened
- Open or commercial
- Open
- Maturity
- PrototypeOur reading, 16 November 2026
- Safety
- S0
- Review
- Listed 16 November 2026by Public Protocol editors
- Checked
- 16 November 2026
- Next check due
- 16 May 2027
- Steward
- Public Protocol editors
Where it sits in the blueprint
- Stages
Learn, Build
- Elements it supports
- Context
Climate, setting, scale, cost and skill: not stated in this entry. Check the source against your own place and numbers.
Verification
No verification is recorded on Public Platform for this entry. Records are added when someone reads it for accuracy, reviews it with a stated qualification, builds it or measures it.
Record
- How it came in
Seed catalogue, 16 November 2026. Listed from the sensors, robotics and AI machines research of 2026-11-16. The source was opened and described on the date shown.
- Licence
CC BY 4.0 (paper); model licence not stated on the page opened. Read at https://arxiv.org/abs/2406.09246. How licences are recorded
- Something wrong?