Bright key facts / Experimental
Teaching a robot from demonstrations
OpenVLA gives robotics teams a pretrained vision-language-action model plus instructions for fine-tuning it on their own robot demonstrations.
- AI’s role
- The model reads an image and language instruction, then predicts actions for a robot manipulator.
- Documented result
- The official release provides checkpoints, LoRA and full fine-tuning paths, configuration files, and evaluation instructions for supported robot environments.
- Important limitation
- Repository code is MIT, while checkpoints inherit Llama 2 Community License restrictions. Laboratory evaluations do not establish safe unattended operation.