DeepMind's Gemini Robotics 2 Controls Whole-Body Humanoid Robots

Gemini Robotics 2 is DeepMind's most ambitious embodiment push yet: a single model family that controls whole-body motion — legs, torso, arms, and fingers — rather than just manipulation. The headline capability is cross-platform generality, running the same underlying model across a range of hardware from two-armed research rigs to Apptronik's Apollo 2 humanoid, and executing everyday tasks like fetching snacks and changing lightbulbs.

The technical significance is the move from tabletop grasping (the prior generation's focus) to full-body coordination and balance, which requires the model to reason about locomotion, dynamics, and multi-limb sequencing simultaneously. Building it as a platform-agnostic controller is a bid to become the 'foundation model for robots' — the layer robot makers license rather than build.

DeepMind announced it alongside the Gemini 3.8 Live voice models, but robotics is a distinct thrust aimed at the humanoid gold rush where Tesla Optimus, Figure, and others are competing. Google's edge is its multimodal model depth; its gap is hardware — hence partnering with Apptronik rather than building its own robot.

Scientific American framed it as Google wanting Gemini to 'power many different robots,' the same horizontal-platform strategy that made Android dominant in phones. The caveats are the usual ones for robotics demos: curated tasks, controlled environments, and no data on reliability, latency, or failure rates in the wild. Watch whether third-party robot makers actually adopt Gemini Robotics 2 as their brain, which is the real test of the platform thesis.