Google DeepMind dropped Gemini Robotics 2, a cross-platform software stack that can control a wide range of different humanoids. The single AI model allows robots to apply real-time reasoning to perform a myriad of complex full-body movements. Video demonstrations show Gemini-powered robots walking, crouching, stretching, and maneuvering objects to clean up a cluttered room, as well as perform tasks like screwing in light bulbs and tying trash bags.

According to the announcement, Gemini Robotics 2 can translate a human-level intent into synchronized full-body control. Google notes that a lightweight version of the AI is capable of running locally on-device, allowing the software to seamlessly adapt to new robotic bodies in just a few hours. 

Google is currently using Apptronik’s Apollo 2 humanoid hardware as its primary native testing platform for full-body integration. Gemini Robotics On-Device, the compact version capable of running locally on a robot’s onboard processing chips, can handle balance, leg movement, and hand coordination from a single prompt. This is the most striking advancement in this model—eliminating the need for a constant, high-speed cloud connection during tasks.

It is able to seamlessly transition from controlling a single tabletop arm to driving full humanoid systems. During an official demo, the Apptronik Apollo 2 was asked to put the watering can into the green bin on the bottom shelf. The robot Apollo not only completes the task but also takes a few steps to the shelves, bends down, and places it back in its place. 

Google claims Gemini Robotics 2 introduces advanced physical dexterity across different hardware end-effectors. The model enables multi-fingered bionic hands to handle delicate actions like tying knots or sealing ziplock bags. It also allows two-finger grapplers to balance and handle tight industrial packing. Furthermore, the upgrade introduces multi-robot collaboration, allowing robots to autonomously communicate and work together. 

Google is advertising Gemini Robotics 2 as a software ecosystem that could dominate physical automation. However, the announcement has not yet disclosed the pricing model. There is no verifiable data to prove how these models will cut operating costs or deliver a realistic return on investment (ROI) as of yet. 

If this becomes quickly scalable and accessible, hardware manufacturers can focus entirely on building better, cheaper physical parts and install Gemini to handle the thinking, walking, spatial balancing, dexterity, and safely navigating real-time environments. Gemini Robotics ER 2 can natively use Google Search, custom Application Programming Interfaces (APIs), and live internet data to solve real-world problems on the go. As detailed in the official announcement, if a robot encounters an unfamiliar object in a facility, it can search the web to diagnose what it is seeing and figure out how to fix it. 

Google noted the importance of safety, stating, ‘’Gemini Robotics 2 specifically advances robotics safety for navigating the uncertainty of the real world and collaborating alongside humans.” To achieve this, Google introduced ASIMOV-Agentic to navigate safety and uncertainty initiated by the Vision-Language-Action (VLA) model. In the event of an anomaly, the reasoning agent overrides the system to maintain the robot’s ability to know its own limits, avoid dangerous shortcuts, and proactively ask for human assistance instead of blindly crashing ahead. 

Robot24.com Business News TAKE:

Google is aiming to become, for robots, what Android is for phones. The company is betting that hardware manufacturers will opt for a proven AI model rather than spending billions building their own, which could take years. However, without transparency in pricing, it is too early to gauge the impact or the shift this will bring to the industry.