DeepMind teams up with Agile for robotics
Google DeepMind is partnering with Agile Robots to marry Agile’s hardware and robotics stack with DeepMind’s Gemini Robotics models so teams can work on deployment, data collection and iterative training in real environments. The alliance signals that DeepMind is investing in embodied, multimodal systems and practical robotics pipelines rather than only paper‑scale benchmarks. For candidates, it points to demand for people who can bridge modelling, hardware constraints and real‑world data collection. (evertiq.com)
Robots have been good at doing the same motion a million times. They have been bad at the ordinary stuff humans do without thinking, like finding the right object in a messy bin or adjusting when a part is slightly out of place. (deepmind.google) That gap is what robotics researchers call embodied intelligence: software that does not just predict words on a screen, but sees a workspace, understands a request, and turns that into movement in the real world. Google DeepMind built Gemini Robotics for exactly that job. (deepmind.google) Google DeepMind says Gemini Robotics is based on Gemini 2.0 and is designed to let robots perceive, reason, use tools, and interact with humans across different robot shapes. In March 2025, the lab also published a paper showing the model could be adapted to new tasks from as few as 100 demonstrations. (deepmind.google) (arxiv.org) The hard part is that robot intelligence does not live in software alone. A model can look impressive in a lab video and still fail on a factory floor where cameras shake, lighting changes, and parts arrive with tiny variations every hour. (arxiv.org) (deepmind.google) That is why this new deal matters. Agile Robots said on March 24, 2026 that it formed a strategic research partnership with Google DeepMind to integrate Gemini Robotics foundation models with Agile’s hardware and robotics stack. (agile-robots.com) Agile is not a tiny lab supplier. The company says it has installed more than 20,000 robotics solutions worldwide, which gives DeepMind something its models need badly: real machines doing real work in real environments. (agile-robots.com) (evertiq.com) The partnership is supposed to run as a loop. Deploy robots, collect data from what they actually see and miss, retrain the models, and send the improved system back out again. (automationworld.com) (evertiq.com) Google DeepMind has been building that stack in pieces. Gemini Robotics-ER 1.5, released to developers in late 2025, handles the high-level reasoning step, like planning the sequence of actions, while Gemini Robotics 1.5 handles the physical execution step. (deepmind.google) (ai.google.dev) Put simply, one model is the foreman and one model is the pair of hands. Agile brings the arms, sensors, controllers, and deployed systems that let those models learn where the plan breaks when metal meets conveyor belt. (deepmind.google) (agile-robots.com) This also fits a wider pattern. TechCrunch reported that Agile is one of several robotics companies partnering with Google DeepMind, which suggests the lab is trying to spread its models across multiple hardware platforms instead of betting on a single robot body. (techcrunch.com) The bet here is that robotics will not be won by the best language model alone. It will be won by whoever can connect models, motors, sensors, safety checks, and constant real-world retraining into one working pipeline. (deepmind.google) (automationworld.com)
