Meta AI Updates: July 28, 2026

1. Meta Vision Models Power Assistive Robotics at the University of Pittsburgh

Meta described how its open vision models are being used in the University of Pittsburgh’s RAMMP (Robotic Assistive Mobility and Manipulation Platform) project, which builds assistive robotics for wheelchair users and people with limited mobility. The system pairs DINOv3, used as a compact and efficient visual foundation, with the Segment Anything Model (SAM) to let robots understand their surroundings, navigate safely, and manipulate objects in response to natural language and image queries. To run on battery-powered edge hardware with tight size and weight constraints, the team reduced memory footprint and used lower precision where appropriate, and fine-tuned an RF-DETR detection model on DINOv2 embeddings with SAM used for auto-labeling. The project, funded by up to 41.5 million dollars from ARPA-H, illustrates general-purpose vision foundations being layered with lightweight task-specific modules for real-time robotics. Source