Google DeepMind has hired Aaron Saunders, the long-time engineering leader and former CTO of Boston Dynamics, in a move that signals a renewed push into advanced robotics. Saunders spent more than two decades at Boston Dynamics, where he played a central role in the development of robots such as Atlas and Spot. His shift to DeepMind underlines the company’s intention to bring its AI models into real-world, physical systems.
DeepMind has recently been exploring how its Gemini models can operate beyond simulations, with the aim of creating a unified software layer for a broad range of robotic platforms. CEO Demis Hassabis has described this goal as developing the equivalent of an “Android for robots” — an operating environment where large AI models can control different types of machines without being tailored for each individual system.
The appointment of Saunders adds extensive hardware expertise to that effort. While DeepMind has a long record in software-driven robotics research, integrating high-capacity AI into reliable physical systems remains a substantial challenge. Issues such as the sim-to-real gap, mechanical variability, and safety constraints continue to limit the deployment of general-purpose robots outside structured environments.
Industry analysts note that the move places DeepMind more directly in competition with companies developing general-purpose humanoid and mobile robots. Firms such as Tesla, Figure AI and Agility Robotics have emphasised the importance of tight integration between AI and hardware. DeepMind’s approach, by contrast, focuses on a model-centric stack that could be applied across different robot formats if the software proves flexible enough.
DeepMind has not disclosed specific details about upcoming hardware projects or robot prototypes. However, the combination of Gemini-based control systems and Saunders’ background in complex mechanical platforms suggests the company intends to accelerate its embodied-AI efforts. How rapidly those ambitions translate into deployable machines will depend on whether DeepMind can bridge the gap between high-performance models in simulation and consistent performance in the physical world.
Photo credit: Google, Iwan Baan
