VLA-Edge: An Edge-Deployed Vision-Language Model Framework for Autonomous Robotic Manipulation via Agentic Reasoning

The deployment of large language models in robotic manipulation has largely relied on cloud-based inference, introducing latency, privacy risks, and network dependency. This paper presents VLA-Edge, a fully offline, edge-deployed framework that runs a 2-billion parameter Vision-Language Model (B01X-2B-Edge-Robotics) natively on Apple Silicon using the MLX framework. At its core is a ReAct-based agentic loop that enables a 6-DOF robotic arm to interpret natural language commands, perceive its workspace via RGB camera, and autonomously execute multi-step pick-and-place tasks. Supporting components include an ArUco marker-based homography pipeline for accurate pixel-to-millimeter transformation and a Three.js digital twin for visualization and safety. Experimental results show end-to-end single-step task completion in under 2.1 seconds while maintaining complete data privacy, zero cloud cost, and full offline operation. This is a preprint. The work is currently under review / submitted to a conference.