The Robot Brain That Moves the Whole Body

Google DeepMind's latest robotics models connect language, vision, planning, and motor control across humanoids and other robot bodies. This collection tests the leap from striking demonstrations to dependable work.

7 sources / Updated August 26, 2026

WebGoogle DeepMindGoogle DeepMindGemini Robotics 2 brings whole-body intelligence to robotsOfficial launch and demo package for whole-body control, dexterity, on-device operation, cross-embodiment adaptation, and multi-robot coordination.Read original
What changed
The model family joins visual understanding and language instructions to direct movement across an entire robot, not just a gripper at a fixed station. DeepMind also presents an on-device version and a reasoning model for planning. The most notable claim is transfer: a shared model can adapt to different robot forms with limited additional data. That shifts the research target from one carefully programmed machine toward a reusable control layer that can serve several bodies and coordinate more than one robot.
WebGoogle DeepMindGoogle DeepMindGemini Robotics ER 2 model cardOfficial model scope, evaluation context, limitations, and links to safety analysis for the embodied reasoning component.Read original
What remains unproved
The public material does not establish reliable, economical performance through long shifts in messy homes, hospitals, or factories. Demonstration selection can hide recovery time, human resets, failed attempts, and narrow environmental assumptions. Cross-body transfer is promising, but deployment also depends on hardware durability, safe force limits, latency, maintenance, and clear failure behavior around people. A model that completes diverse clips is not yet a worker that can meet a service-level target without close supervision.
WebarXivarXivGemini Robotics: Bringing AI into the Physical WorldTechnical foundation for the earlier Gemini Robotics family, including few-shot task learning and adaptation to new robot embodiments.Read original
What to watch
Watch for independent task suites, full uncut runs, recovery from mistakes, and results reported by robot body and environment. The useful benchmarks will include completion rate, intervention rate, energy use, cycle time, and safety stops, not only semantic understanding. Also track whether the on-device model keeps working when network access fails and whether one policy can move among bodies without extensive retraining. Commercial pilots with disclosed work hours would mark a stronger step than another polished capability reel.
WebarXivarXivEvaluating Gemini Robotics Policies in a Veo World SimulatorResearch on using generated video worlds to probe nominal behavior, unusual conditions, generalization, and physical or semantic safety.Read original

Loading collection…

What changed

What remains unproved

What to watch