Google DeepMind's Running Guide agent navigates blind and low-vision runners outdoors using on-device Gemma 4

Google DeepMind published details on May 20, 2026 of an accessibility agent that guides blind and low-vision athletes while running outdoors in real time, using a Pixel 10 Pro smartphone worn on the chest and Gemma 4 for multimodal scene understanding.

What's new

The Running Guide agent is designed for blind and low-vision (BLV) athletes. As DeepMind notes, "for blind and low-vision athletes, running has traditionally required a physical tether — whether it's a human guide or a painted track line." The system replaces that physical dependency with an AI agent delivering continuous audio navigation.

Architecture details:

• A Pixel 10 Pro is worn chest-mounted, facing forward
• On-device segmentation models run fully offline for obstacle detection, maintaining function without a network connection
• Gemma 4 E4B (a quantized variant suited for mobile deployment) handles multimodal scene understanding
• Smarter Frame Selection processes only high-entropy camera frames rather than every frame, reducing compute and improving latency
• Three collaborative agents: a Planner (route planning), a Coach (real-time alerts and steering cues), and a Break agent (managing rest intervals)

The system delivers "immediate 'STOP' alerts and steering cues — heard as directional ticking sounds — so runners maintain a reliable sense of direction."

Context

Running outdoors with a physical guide runner has logistical constraints: it requires a sighted partner, coordination around pace and schedule, and in competitive settings, a registered guide. Many BLV athletes want to train independently. Track running with a painted guide line addresses the independence issue but limits athletes to controlled facilities.

This DeepMind project builds on several converging capabilities: efficient on-device inference via Gemma 4's compact architecture, multimodal scene understanding that can interpret visual context in real time, and agentic systems that decompose a complex navigation task into discrete specialized roles (planning, real-time coaching, rest management).

The multi-agent decomposition — Planner, Coach, Break — mirrors patterns from production AI systems where specialized subagents handle discrete subtasks rather than a single model managing all decision-making. Applying this architecture to a real-time safety-critical physical task is a meaningful step.

Why it matters

The Running Guide agent demonstrates that agentic AI is moving beyond software productivity into physical-world accessibility applications with safety-critical latency requirements. Delivering reliable real-time obstacle detection and directional audio without a network dependency is a harder problem than most enterprise AI applications — the consequences of a missed obstacle are immediately physical.

The choice to run segmentation models fully on-device is an important design constraint: it means the system works in environments with poor connectivity and avoids the latency of round-trip cloud inference. This is a practical model for other real-time AI assistance systems where reliability outweighs the appeal of the latest frontier model.

For Google DeepMind, publishing this research demonstrates that Gemma 4's efficiency is sufficient for safety-critical mobile deployment, not just benchmark performance. Whether Google plans a consumer product based on this work has not been announced; the publication frames it as a research demonstration.