Gemini Robotics 2 gives humanoid robots full-body control
SkimNews Take
Unified whole-body control — coordinating balance, locomotion, and dexterity in one model — has been the main bottleneck for general-purpose humanoids; achieving it on-device signals a shift from lab demos toward field-deployable systems.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Google DeepMind announced Gemini Robotics 2, expanding control from the upper body of a humanoid robot to its full body, enabling movements from feet to fingertips.
- The model supports whole-body actions — walking, crouching, stretching, and object manipulation — demonstrated on Apptronik's Apollo 2 robot bending to pick up a watering can and retrieving specific items from a shelf.
- Gemini Robotics 2 adds dexterity through five-fingered-hand control, letting robots perform tasks like sealing a Ziploc bag, tying a trash bag, and unscrewing a lightbulb.
- The companion Gemini Robotics ER 2 embodied-reasoning model was upgraded for longer multi-step tasks, recognition of when tasks begin and end, and is described by Google DeepMind as its "safest robotics model to date" with better human-detection that can trigger safety stops.
- Google showed multi-robot collaboration in which an Apollo 2 humanoid instructed Google's dual-arm robot to place tools in a bin while cleaning a garage.
- The improved Gemini Robotics On-Device Model can run without an internet connection and now adapts faster to new robot embodiments with different shapes, sensors, and degrees of freedom.
- Google DeepMind acknowledged that the robots "have more to advance in movement speed," framing the update as a step toward more complex, real-world tasks requiring whole-body coordination.
Why it matters: By upgrading from upper-body to full-body control and adding multi-robot collaboration, Google DeepMind gives humanoid-robot hardware partners like Apptronik a software path to more practical home and workplace tasks — but the company itself concedes movement speed still lags, tempering the near-term real-world utility of these demos.
