Google DeepMind’s Gemini Robotics 2 can control humanoid robots from feet to fingertips, report says

RoboZaps reported on July 31, 2026, that Google DeepMind’s “Gemini Robotics 2” can control humanoid robots from feet to fingertips. The article presents the model name and scope of control as its main facts. The report does not provide details about the conditions under which performance was evaluated or the specific robot platforms involved.

A report that Google DeepMind has outlined a scope approaching full-body control of humanoid robots gives robotics control models a more concrete benchmark: the range of movements they can handle. This can help developers identify what they need to verify as they distinguish research demonstrations from real-world operation.

What is new in this report is its framing of Google DeepMind’s robotics model as covering a scope close to full-body control of humanoid robots, rather than just perception or conversation. However, based on the excerpt, this is headline information about the model’s scope of application, not a definitive performance claim. The phrase “from feet to fingertips” suggests integrated control that includes posture maintenance and locomotion as well as upper-body manipulation, but it is unclear which robot platform was used and under what conditions. A related article in this newspaper, Dyna Robotics executive discusses “field deployment,” humanoid.guide reports, focused on deployment design outside the laboratory. Viewed alongside this report, the focus in robotics appears to be shifting from “whether it can be done” to “what range of movements can be controlled reliably.” However, this article alone does not confirm comparative metrics needed for field deployment, such as repeatability, safety, the breadth of training data or failure rates. One potential industry implication is that as control models cover a broader range, models, training data and evaluation environments may carry more weight relative to the robot hardware itself. This report, however, does not specify the model’s training infrastructure or supported platforms, and it does not establish an advantage over competing models. The next points to verify are which tasks it can perform and with what degree of reproducibility, the conditions of real-world demonstrations, and whether it can be extended to platforms other than humanoid robots.

For Japanese robotics companies and research institutions, the scope of control delegated to models in humanoid robots will be a point of comparison. However, the report does not detail supported platforms or training conditions, limiting its direct implications for Japan.