Google DeepMind's Gemini Robotics 1.5 Brings Agentic AI to Physical Machines

Google DeepMind unveiled Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 on September 25, 2026 — two new AI models that add agentic capabilities to robots, enabling multi-step reasoning, tool use, and generalization across physical tasks.

Carolina Parada, DeepMind's senior director and head of robotics, described the release as an "important milestone" toward artificial general intelligence in the physical world.

Two Complementary Models

Gemini Robotics 1.5 is a vision-language-action (VLA) model that converts visual information and natural language instructions into motor commands. Show a robot a scene, tell it what to do, and the model generates the physical actions required.

Gemini Robotics-ER 1.5 is a vision-language model specialized in understanding physical spaces. It creates multi-step plans to complete tasks and can natively call tools — including Google Search for information lookup and user-defined functions for custom integrations.

The ER model is available to developers through the Gemini API in Google AI Studio. The VLA model is currently available to select partners, with broader access planned.

Beyond Reactive Commands

Previous robotics AI systems largely reacted to immediate commands: pick up this object, move to that position. The new models introduce genuine planning — breaking complex tasks into steps, adapting when conditions change, and seeking external information when their internal knowledge is insufficient.

A robot tasked with preparing a workspace might search for the correct tool specifications, plan an efficient movement sequence, execute the physical actions, and adjust if an obstacle appears — all without human intervention at each step.

Parada emphasized that the models can generalize across tasks and transfer learning from one physical context to another, reducing the need for task-specific training data.

Demonstrations and Partners

Google DeepMind demonstrated the models with Apptronik's APOLLO humanoid robot, showing coordinated movement and object manipulation guided by the new AI systems. Select partners are already integrating Gemini Robotics 1.5 into commercial robotics platforms.

Implications for Industry

Manufacturing, logistics, healthcare, and home robotics have long promised AI-powered automation. The bottleneck has been the gap between language understanding and physical dexterity — robots that could talk but not act, or act but not reason.

Agentic robotics models bridge that gap. A warehouse robot that can reason about inventory placement, search for handling instructions, and physically execute complex packing sequences represents a qualitatively different capability than scripted automation.

Challenges Ahead

Physical AI introduces safety considerations that software-only agents avoid. A coding agent that makes a mistake produces a bug. A robotics agent that makes a mistake can cause physical harm. Permission systems, hardware interlocks, and human oversight remain essential even as models become more autonomous.

Sim-to-real transfer — the gap between training environments and physical reality — continues to challenge robotics AI. DeepMind's demonstrations are promising, but scaling to diverse real-world conditions remains an open engineering problem.

The AGI Connection

DeepMind explicitly frames robotics as a path to AGI. Physical intelligence — understanding and manipulating the three-dimensional world — is a capability that current language models lack entirely. By building models that connect language reasoning to physical action, DeepMind is pursuing a form of intelligence that cannot be achieved through text alone.

Whether Gemini Robotics 1.5 represents a meaningful step toward that goal or an impressive but narrow capability will become clearer as developers and partners push the models beyond controlled demonstrations.