“We are building the world’s most capable humanoid,” said Alberto Rodriguez, director of robot behaviour for Atlas at Boston Dynamics, in a recent unveiling. That ambition today takes on quite another dimension, with Boston Dynamics integrating Google DeepMind’s Gemini Robotics AI foundation models into its latest generation Atlas humanoids and Spot quadrupeds, commencing trials inside Hyundai’s automotive manufacturing facilities.
Boston Dynamics has long been known for running, jumping, and running complex terrains; this commercial launch of Atlas, however, showed a shift toward enterprise-grade deployment. The fully electric humanoid is designed for industrial versatility-from lifting as heavy as 50 kg to adapting to dynamic environments and working independently for extended periods. It can re-charge itself without human intervention, connect directly to manufacturing execution systems, and replicate learned tasks across an entire fleet. Such capabilities are being tested in Hyundai’s Robotics Metaplant Application Center, where production lines are already heavily automated and optimized for robotic integration.
A partnership with Google DeepMind aims to give Atlas its cognitive upgrade. Gemini robotics models rely on large-scale multimodal AI systems that will enable robots to perceive, reason about, and communicate with humans using natural language. That would mean Atlas understanding complex requests from users, splitting goals into smaller tasks, and changing paths when unexpected barriers appear. The aim is to bring AI into the physical world, and explore what can be done when advanced reasoning meets advanced mobility, said Carolina Parada, DeepMind’s senior director of robotics.
The trial is part of a wider robotics strategy at Hyundai, Boston Dynamics’ majority shareholder since 2021. It had promised to deploy tens of thousands of robots across its manufacturing network as part of a new facility capable of producing 30,000 robots annually. Starting in automotive plants-where precision, repeatability, and safety standards are already high-the combination of humanoids and quadrupeds could be tested out in a controlled, high-throughput setting.
Safety is an important factor: collaborative robots such as Atlas must work within tight guidelines, such as those detailed in ISO/TS 15066, that dictate limits on the level of force and pressure tolerable upon contacting a human along with acceptable power and speed limits. This keeps accidental contact from becoming injury. Atlas has the incorporation of human detection systems, fenceless guarding, and integration with barcode or RFID workflows to meet these requirements without losing productivity.
Results from data gathered during the Hyundai trials will be crucial. Real-world deployments generate rich sensory inputs-from camera and LiDAR readings through to force feedback and task-level telemetry that can reveal where robots succeed and where they deviate from expected performance. This feedback loop lets engineers refine AI models, address edge cases, and build more robust control systems. As robots encounter variations in lighting, workspace clutter, or object shapes, those experiences become part of a growing dataset that strengthens future behavior across the fleet.
Boston Dynamics has already demonstrated the capability of large AI models in robotics through various collaborations, including work with the Toyota Research Institute in which it had Atlas execute long sequences of complex manipulation and locomotion tasks under one neural control system. Integration with Gemini Robotics aims at extending this capability whereby robots can think before doing something, explain their reasoning, and adjust midtask without human intervention.
The Hyundai trials are only a beginning, but they represent an intersection of mechanical engineering, AI research, and industrial automation. If the experiments pan out, lessons learned could inform deployments in logistics and construction and other sectors where adaptable, human-scale robots could work alongside people. For now, Atlas and Spot will have to navigate the structured yet demanding environment of automotive manufacturing-testing not only their mechanical limits but also the depth of their decision-making intelligence.