Introduced May 20, 2026, The Running Guide Agent is an on‑device accessibility system that pairs low‑latency edge

On May 20, 2026 DeepMind introduced the Running Guide agent, an accessibility system designed to let blind and low‑vision (BLV) athletes run without a human guide or painted track lines. The project reframes assistive running from simple path‑following toward real‑time spatial reasoning, with the stated goal of enabling unassisted independence for runners. Visual input comes from a chest‑mounted Pixel 10 Pro, and guidance is delivered through concise audio prompts.

The system uses a hybrid, dual‑path architecture that separates immediate safety from higher‑level scene understanding. An on‑device segmentation model runs at the edge to provide zero‑latency steering cues and emergency alerts, while a Gemma 4 E4B‑powered reasoning path handles complex interpretation of the environment. To keep responsiveness during fast activity, the Gemma 4 E4B path applies “Smarter Frame Selection,” processing only high‑entropy frames — moments with sudden terrain changes or new obstacles — instead of every video frame. That selective processing is intended to prioritize the most relevant inputs and deliver faster, more actionable coaching while minimizing computational load.

Running Guide is organized as a multi‑agent framework. A Planner agent uses Gemma 4 function calling to fetch weather and Google Maps data, discuss goals with runners and calibrate a digital starting line. A Coach agent issues brief verbal alerts triaged into DANGER (immediate evasive action), WARNING (nearby runners or obstacles) and NOTICE (upcoming curves). A Break agent schedules rest intervals and manages pause/resume.

Beyond the chest‑mounted phone, the team is prototyping intelligent eyewear that streams a wider, steadier field of view to the Pixel device. The eyewear is intended to improve coverage and stability of visual input, combining wearable sensors with on‑device inference to optimize the data fed to the multimodal models.

DeepMind is pairing engineering teams with BLV runners through a partnership with SG Enable (Singapore’s focal agency for disability and inclusion) to test the system in real‑world conditions and iterate on design. For builders of safety‑critical mobile agents, the project demonstrates a practical pattern: combine zero‑latency edge segmentation, selective frame processing and on‑device reasoning to meet responsiveness and trust requirements for high‑speed activities.