A vision-language-action model that runs on-device
Inference moves off the datacentre and onto an embedded module drawing under 30 watts.
The team distilled a 7B parameter policy down to something that fits in 8GB.
Trade-offs
Success rates drop roughly four points against the full model.