Google DeepMind Brings Project Astra to ICML as a Real-World Phone Agent

Project Astra is Google DeepMind's most ambitious bet on what an AI assistant could actually become. Not a chatbot, not a voice interface, but a persistent, multimodal agent that watches the world through your camera, listens to your environment, remembers your past, and takes actions on your behalf. This week, Google Research brought it to ICML 2026 in Seoul, giving the machine learning community its first hands-on look at the prototype in a research setting.

What Google showed at ICML

Google invited attendees at ICML 2026 to stop by Booth B206 to explore their latest advancements in machine learning, particularly across computer vision and machine perception. The Project Astra session was one of the most prominent on the schedule. The demo, titled "Project Astra: On the Way to Building a Universal AI Assistant," was presented by Yana Lunts, Sam Holt, Pavel Dubov, and Stefan Moser. The same team ran both a live Q&A and an interactive kiosk session, giving researchers multiple opportunities to engage with the prototype directly.

What Astra actually is

Project Astra is a research prototype exploring breakthrough capabilities for Google products, on the way to building a universal AI assistant. The key word here is agent, not assistant. Astra sees through the camera, controls your phone, opens apps, makes phone calls, reads emails, and remembers your preferences over time. The distinction from something like Gemini Live is meaningful: Gemini is the AI model, Astra is the agent built on top of it. Gemini Live enables voice conversations. Astra goes further: it controls the phone, uses Google tools, remembers context across sessions and devices, and can act proactively.

The technical pillars

The technical foundation of Project Astra represents a departure from the "token-in, token-out" architecture of early large language models. To achieve fluid, human-like responsiveness, Google DeepMind engineers focused on three core pillars: multimodal synchronicity, sub-300ms latency, and persistent temporal memory. These are not incremental improvements to a chat interface. They are the engineering prerequisites for an agent that can participate in the physical world.