Same Open-Source Vision Models, Two Very Different Deployments

Meta’s open-source computer vision models, DINO and Segment Anything Model (SAM), are now running inside a $41.5 million ARPA-H-funded robotics program at the University of Pittsburgh’s Human Engineering Research Laboratories. The project, RAMMP (Robotic Assistive Mobility and Manipulation Platform), aims to give wheelchair users robotic systems that understand natural language and environmental context — a real problem at real scale, given the 5.5 million wheelchair users in the US and more than 100,000 wheelchair-related injuries treated in US emergency departments each year. DINO handles scene understanding as a self-supervised vision transformer; SAM identifies and outlines objects from minimal prompting. The Pittsburgh team is optimizing both for edge devices, trading recognition precision against power draw and battery life on compact hardware, with partners including Carnegie Mellon, Cornell, Northeastern, and Purdue.

The same model family is doing comparably serious work elsewhere. Meta’s SAM 3 and DINOv3 also power SYNAPS-I, a Department of Energy Genesis Mission project running across 300 NVIDIA A100 GPUs at the National Energy Research Scientific Computing Center. There, the same segmentation-plus-scene-understanding pipeline turns raw imaging data — DOE facilities generate tens of petabytes a year, with some detectors capturing 100,000 images per second — into labeled 3D volumes in about 15 minutes. A demonstration on grapevine drought-resilience research used the pipeline to identify xylem vessels in micro-CT scans, compressing a month of manual analysis into roughly 15 minutes.

What connects a wheelchair-assist robot and a national laboratory’s imaging pipeline isn’t the application, it’s the licensing. Because DINO and SAM are fully open source, both the Pittsburgh team and DOE’s air-gapped national-lab environments can download and fine-tune the same underlying models entirely on their own infrastructure, without routing sensitive research or accessibility data through an external commercial API. That’s the actual argument for open weights that policy debates tend to skip past: not ideology, but who gets to build safety-critical systems without asking a vendor’s permission first.