Agile Robots integrates Google DeepMind foundation models into its robots

German robotics company Agile Robots has joined the growing list of Google DeepMind partners. Under the agreement, Agile Robots will integrate DeepMind's foundation models into its robots, while in return it will collect real-world data on interactions with the physical environment for the research laboratory. Agile Robots was founded in 2018 as a spin-off of the German Aerospace Center (DLR).

The company specializes in manipulator robots with intelligent control — its flagship product Diana is capable of performing precise assembly operations and object manipulation. The headquarters is in Munich, but the company is actively expanding in the Chinese and European markets. For Google DeepMind, this partnership continues a series of strategic agreements with robotics companies.

Previously, DeepMind concluded similar deals with Apptronik, Enchanted Tools, and several other startups. The essence of the scheme in all cases is the same: DeepMind provides access to foundation models that allow robots to master new tasks faster and with fewer examples, while partners in return supply a stream of real-world data about work in the physical environment. Foundation models for robots are large pre-trained neural networks that, by analogy with language models, are capable of generalizing knowledge across different tasks.

Instead of training from scratch for each operation, such a model transfers skills: basic knowledge about the physics of object grasping will help the robot master the assembly of a new part faster. Google DeepMind invests in this direction through the RT-2 project and its successors, which combine computer vision, language understanding, and manipulator control. The key problem that such partnerships solve is a data deficit.

Language models were trained on trillions of tokens from the internet. Robots don't have such a resource: each movement cycle must be physically executed in the real world. That's exactly why the deal structure looks this way: DeepMind needs real movement trajectories, grasping forces, data on failed attempts — the kind of data that Agile Robots' robots generate during normal operation.

For DeepMind, this is a valuable training signal; for Agile Robots, it's access to cutting-edge AI models without the need to independently build research infrastructure. The partnership fits into a broader competitive landscape. OpenAI invested in Figure AI and 1X.

Amazon supports Agility Robotics. Microsoft works with Apptronik. Nvidia is building the Isaac platform and investing in a wide range of robotics startups.

The competition for whose AI platform will become the standard for the next generation of industrial robots is unfolding right now, and each new partnership is another brick in the corresponding ecosystem. For DeepMind, real-world data is critically important also because the so-called sim-to-real gap — the gap between the model's capabilities in simulation and in a physical environment — still remains a serious obstacle. Data from manufacturing partners allows closing this gap significantly faster than within a closed laboratory.

Agile Robots becomes part of a forming ecosystem in which several major AI labs are essentially building parallel standards for intelligent robots. Which of these standards will win — or whether several will survive — will largely determine the architecture of industrial automation for the next decade.