Google DeepMind has launched a new Running Guide agent that uses artificial intelligence to help blind and low-vision athletes run without physical guides or tethers, marking a significant shift from traditional accessibility approaches. The technology combines on-device processing with advanced multimodal AI reasoning to deliver real-time directional feedback through auditory cues, allowing runners to maintain independence and safety at high speeds.

How Does the Running Guide Agent Work?

The Running Guide agent uses a chest-mounted Pixel 10 Pro smartphone to view the path ahead and guide the user through directional ticking sounds and verbal alerts. The system employs a hybrid architecture designed specifically for high-speed activities where safety is critical. Rather than relying on cloud processing, the technology runs primarily on the phone's custom silicon to ensure ultra-low latency, meaning the system responds almost instantly without needing an internet connection.

The system combines two complementary processing approaches:

• On-Device Segmentation: A lightweight model runs entirely offline on the Pixel 10's processor, delivering immediate "STOP" alerts and steering cues to keep runners safe even without cellular connectivity.
• Advanced Reasoning with Gemma 4: The system leverages Gemma 4 E4B, Google's latest reasoning model, to process complex multimodal inputs including images and text for high-level scene understanding, all processed directly on the device.
• Smart Frame Selection: Instead of analyzing every camera frame, the AI only examines "high-entropy" frames where significant changes occur, such as sudden terrain shifts or new obstacles, reducing processing demands while maintaining accuracy.

What Makes This Different From Previous Accessibility Tools?

The Running Guide agent represents a substantial evolution from Google DeepMind's earlier Project Guideline work. While previous solutions required painted track lines or human guides physically tethered to the runner, this new system enables truly independent running through spatial reasoning and environmental understanding. The technology uses a collaborative multi-agent framework where different AI agents handle specific tasks during the running experience.

The system includes a planner agent that uses function calling to gather weather data and Google Maps information while chatting with the runner to establish workout goals. A coach agent operates during the run to deliver concise verbal alerts, triaging feedback into a strict hierarchy: "DANGER" for immediate evasive action, "WARNING" for nearby obstacles or runners, and "NOTICE" for upcoming track curves. A break agent manages rest intervals, allowing athletes to pause and resume their session at any time.

Where Is This Technology Being Tested?

Google DeepMind is partnering with SG Enable, Singapore's focal agency for disability and inclusion, to test and refine the Running Guide agent for real-world use. This collaboration ensures the technology meets the actual needs of blind and low-vision runners through direct engagement with the community. The partnership is part of a broader expanded collaboration between Google and Singapore's government to deploy frontier AI for public good across multiple sectors including healthcare, education, and scientific research.

The company is also prototyping the Running Guide agent on intelligent eyewear, which could provide a wider and steadier field of view compared to a chest-mounted phone. Wearable glasses would stream directly to a Pixel device, seamlessly blending hardware and ambient AI to optimize the data fed to the multimodal models.

"This partnership builds on years of close collaboration with Google, and we are pleased to take it to the next level. Bringing frontier AI into our public services and enterprises is central to Singapore's AI ambitions," said Chng Kai Fong, permanent secretary for Digital Development and Information.

Chng Kai Fong, Permanent Secretary for Digital Development and Information, Singapore Ministry of Digital Development and Information

How Does This Fit Into Google DeepMind's Broader AI Strategy?

The Running Guide agent exemplifies Google DeepMind's focus on deploying frontier AI models to solve real-world problems. The technology demonstrates how advanced reasoning capabilities, like those in Gemma 4, can be adapted for accessibility applications that directly improve quality of life. The Singapore partnership also includes other accessibility initiatives, such as a Gemma-powered running assistant for blind and low-vision athletes, alongside broader efforts in healthcare, education, and enterprise innovation.

Google DeepMind's presence in Singapore as part of its global National Partnerships for AI initiative reflects the company's commitment to responsible AI deployment. The partnership includes collaboration with Singapore's National Research Foundation to train local researchers on agentic AI tools for science, as well as work with the Agency for Science, Technology and Research to accelerate research translation in materials and life sciences.

"Through this expanded partnership with the Singapore Government, we are putting AI into action by combining the best of our technology, R&D expertise and local talent to accelerate AI for the public good. This also creates a scalable blueprint for responsible AI innovation, built in Singapore for the world," said Ben King, country managing director of Google Singapore.

Ben King, Country Managing Director, Google Singapore

The Running Guide agent represents a meaningful step toward unassisted independence for athletes with vision impairments. By combining on-device processing for safety with advanced multimodal reasoning for scene understanding, Google DeepMind has created a system that prioritizes both reliability and capability. As the technology moves from testing with SG Enable toward broader deployment, it demonstrates how frontier AI models can be adapted to solve accessibility challenges that traditional technology has struggled to address.