Summary: Google DeepMind unveiled Gemini Robotics 2 on July 30, 2026 with whole-body control for humanoids and bi-arm robots, plus an ER 2 reasoning model and an On-Device 2 variant. The system can drive dexterous hands, coordinate multiple robots, and adapt a new embodiment in about a few hours with fewer than 200 examples. DeepMind also introduced the ASIMOV-Agentic safety benchmark.
On July 30, 2026, Google DeepMind unveiled Gemini Robotics 2, carrying its vision-language-action line one step further: robot motion is no longer confined to the upper body but covers the whole body, from feet to fingertips. Carolina Parada, who heads robotics at DeepMind, frames this generation in the announcement as “the intelligence layer powering the next generation of truly adaptable robots.”
A Single System, Three Models
Three models make up Gemini Robotics 2. The flagship is Gemini Robotics 2 itself — the strongest vision-language-action model (VLA), which turns visual and language input straight into motor control and can fully drive humanoids as well as bi-arm robots. Gemini Robotics ER 2 is the embodied reasoning model (VLM), serving as the robot’s high-level brain: it talks with people, grasps the physical world, and plans multi-step tasks that stretch across several minutes. Gemini Robotics On-Device 2 puts efficiency first — a VLA designed to run offline on the robot’s own hardware. On availability, ER 2 is live on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while the remaining two models open to early-access partners first.
Whole-Body Control and Dexterous Hands
Whole-body work is where the demo lands hardest. Ask Apptronik’s Apollo 2 humanoid to “put the watering can into the green bin in the bottom shelf” and it heads to the table on its own, lifts the can, walks over to the shelving, and sets it down — earlier generations handled only tabletop jobs with the upper body. As for dexterity, one and the same model checkpoint runs the 22-degree-of-freedom SharpaWave five-finger hand mounted on Apollo 2 through fine actions such as tying knots or sealing a ziplock bag, and on a Franka Duo platform it works standard two-finger grippers for tight packing. DeepMind stays candid about where the limits sit: multi-finger dexterous manipulation is still a hard problem, and movement speed has room left to improve.
Agentic Reasoning and Robots Working as a Team
ER 2 carries itself more like an agent. It surveys the space, works out the steps, hands execution to the VLA, and keeps tabs on progress, correcting itself whenever a step falls through. The sequences this generation can carry out reliably run several minutes and involve hundreds of decisions, and the model also grasps where a task starts and stops, marking the moment key events occur. A new multi-robot collaboration feature lets robots of different types talk to one another and divide up workflows beyond what any single machine could complete; in the cluttered-room demonstration, robots join forces to finish faster.
On-Device 2: A New Body in Hours
Plenty of robotics applications have no network to lean on. On-Device 2 carries forward the “motion transfer” techniques from Gemini Robotics 1.5 and supports multiple embodiments natively: hand it a new robot whose shape, sensors, and degrees of freedom differ sharply, and adaptation usually takes just a few hours and fewer than 200 examples, with the Dexmate, SO101, and Trossen platforms used as demonstrations. For hardware startups, the price of “buy the model, fit the body” falls another notch.
Safety Design: The ASIMOV-Agentic Benchmark
Stronger capability pulls stronger safety testing along with it. DeepMind put forward ASIMOV-Agentic, a benchmark for agentic safety orchestration and uncertainty resolution. It gauges three things: whether the embodied reasoning agent turns down unsafe tool calls issued by a VLA, whether it anticipates if a task can be done at all, and whether it asks a human to step in when it is unsure. ER 2 also stands as DeepMind’s safest robotics model to date on the safety-constraint-following and human-proximity benchmarks: it notices when people come close, fires off safety tool calls, and brings the robot to a safe stop — a key requirement written into collaborative safety standards. A safety technical report released alongside the announcement holds the details.
What It Means for Robotics Developers
Three signals deserve attention. First, the layered arrangement in which an ER model plans and a VLA executes is settling in as the standard architecture for robots, echoing the planner-and-tools division of labor in software agents, and developers can evaluate each layer on its own. Second, with multi-embodiment adaptation down to a few hours and roughly 200 examples, robotics startups can focus their effort on hardware design instead of training models from zero. Third, by placing ER 2 in AI Studio and an enterprise agent platform, Google draws physical AI into the ordinary developer toolchain. DeepMind states its destination plainly — solving AGI in the physical world. That destination is far off, yet this release shifts humanoid robots from demo footage toward tasks people can actually use.