Agile Robots signs Google DeepMind research pact to run Gemini Robotics models on its industrial robots, widening Google's hardware bench

Agile Robots SE, the Munich automation group set up by researchers from the German Aerospace Center (DLR), said on 24 March 2026 that it had formed a strategic research partnership with Google DeepMind to run the Gemini Robotics family of foundation models on its robots. According to the company, the work will begin with industrial use cases in manufacturing, where customers need automation that stays reliable while adapting to changing tasks.

The release is short and leaves most commercial questions open. It discloses no financial terms, no equity stake and no exclusivity, and it does not say which Gemini Robotics model versions will be integrated or which Agile Robots machines will carry them first. What it does give is scale on the hardware side. Agile Robots says it has installed more than 20,000 robotic solutions worldwide; in February 2026 it put its team at more than 1,000 robotics and AI specialists, and in November 2025 it said revenue had doubled every year since its founding to reach €200 million.

Both companies describe the partnership as a data loop. Agile Robots contributes its industrial platform and live deployments, data from joint testing and real operation is fed back to improve the models, and better models are meant to let the robots take on more tasks in more plants. Zhaopeng Chen, Agile Robots' founder and chief executive, said putting the Gemini Robotics models into the company's systems placed it at the cutting edge. Carolina Parada, senior director and head of robotics at Google DeepMind, called the agreement an important step in carrying AI's impact into the physical world.

Agile Robots was founded in 2018. Its last equity round confirmed on a primary page in MEKAWIRE data is a $220 million Series C led by SoftBank Vision Fund 2, announced in September 2021. In November 2025 the company launched Agile ONE, its first humanoid for industrial work, with hands that carry fingertip sensors and force and torque sensing in every joint, and a layered AI design that separates task planning, fast reactions and fine motor control. It said full production would start in early 2026 at a new plant in Bavaria. On 4 February 2026 an Agile ONE switched on the Industrial AI Cloud that Deutsche Telekom and NVIDIA built in Munich, where Agile Robots, one of the first customers, said it would train its own foundation models.

The tie with Google is older than the announcement. When Google DeepMind introduced Gemini Robotics on 12 March 2025, it named Agile Robots as one of four trusted testers of the embodied reasoning model Gemini Robotics-ER, together with Agility Robotics, Boston Dynamics and Enchanted Tools, and named Apptronik as its humanoid partner. The March 2026 agreement turns Agile Robots from an early tester into a research partner.

It also completes a busy quarter for Google's robotics effort. On 5 January at CES, Boston Dynamics said it would bring Gemini Robotics models into its Atlas humanoid, with joint research due to start within months and an initial focus on car manufacturing. On 25 February Alphabet moved Intrinsic, the industrial robot software unit it created as an Other Bet in 2021, into Google as a distinct group working closely with Google DeepMind, Gemini and Google Cloud. On 11 February Apptronik, whose Apollo humanoid is one of the test platforms for Gemini Robotics 1.5, closed a $520 million extension that took its Series A past $935 million.

Rival model builders are taking other routes. NVIDIA distributes its Isaac GR00T humanoid models openly to any hardware maker, Figure develops its Helix models in house for its own robots, and Physical Intelligence publishes its π model research and on 3 March released a memory system that lets its vision-language-action models handle tasks longer than ten minutes. Google's approach sits between these, keeping its most capable action model with selected partners while offering the planning model to all developers.

The technology behind the deal was set out with Gemini Robotics 1.5 in September 2025. It pairs two models. Gemini Robotics-ER 1.5, a vision-language model, plans multi-step jobs and can call digital tools such as Google Search; Gemini Robotics 1.5, a vision-language-action model, turns camera images and instructions into motor commands and reasons through a step before it moves. Google DeepMind reported that tasks shown only to its ALOHA 2 research robot during training also worked on Apptronik's Apollo and on a Franka two-arm system. ER 1.5 is open through the Gemini API, while the action model is limited to selected partners, which is why a formal partnership matters to a hardware maker.

For Google, Agile Robots adds a European industrial installed base and a broad portfolio of industrial systems rather than another humanoid alone. For Agile Robots, the agreement offers a frontier model family that a mid-sized hardware company would struggle to train by itself, even as it trains its own models on the Munich cloud. The layered design it described for Agile ONE leaves room for an external planning model above in-house motor control, so the two tracks need not collide.

The open questions are practical. The release gives no timetable, no named customer and no performance target, and industrial buyers judge robots on cycle time, uptime and safety certification, measures on which no Gemini Robotics model has published factory results. The data loop also raises an ownership issue that German manufacturers will press on, namely whether data gathered at their sites can train models that Google later serves to their competitors.

The next markers are concrete. Agile Robots has said Agile ONE production starts in Bavaria in early 2026, Boston Dynamics said in January that joint research with Google DeepMind would start within months, and Hannover Messe in April is the first large industrial fair since the announcement where a joint demonstration could appear.