Google DeepMind released Gemini Robotics ER 2 on July 30, an embodied reasoning model that plans multi-step physical tasks for robots without generating the low-level motor commands itself. The model hands execution to a separate vision-language-action model, a split DeepMind says lets one reasoning layer run tabletop arms, quadrupeds, and humanoids alike. That division of labor determines what a developer can actually build on top of it.
Embodied reasoning, the ER in the name, means spatial understanding and planning about the physical world: tracking task progress from video, judging when a step is finished, and deciding what happens next. It is not a policy that outputs joint torques or wheel speeds. Developers declare a VLA model or a navigation API as a callable tool, then stream video, audio, or text into Gemini Robotics ER 2, which orchestrates those tools rather than driving the actuators itself. The model can also call outside tools such as Google Search.
DeepMind says the new version watches continuous video instead of single frames, so a robot can track its own progress and retry a failed step without restarting the whole task. In one evaluation, the company graded task completion across five bands, from 0 to 20 percent up through 80 to 100 percent, and reported 57.4 percent accuracy, ahead of the prior generation and of competing frontier models it tested against. A second benchmark, moment finding, asks the model to pinpoint the exact video frame where an event occurs, such as when to stop pouring a liquid. Gemini Robotics ER 2 scored 91.3 percent accuracy with a mean error of 0.96 seconds, which DeepMind says it delivers at four times the execution speed of larger rival models and at a fraction of their compute cost. Google ran every one of these evaluations itself; the release includes no independently verified figures.
The model also introduces multi-robot collaboration, letting different machines share a common understanding of a task and hand off pieces of it to each other. DeepMind’s demo pairs Apptronik’s Apollo 2 humanoid with Franka’s F3 Duo arm, working the same shared workflow. A separate demo runs the model against Boston Dynamics’ Spot, using it to orchestrate Spot’s navigation and manipulator APIs so the quadruped can fetch an object on a spoken request.
On safety, DeepMind reports gains on two internal benchmarks it calls Safety Instruction Following and Human Proximity, which test whether the model respects physical constraints and stays aware of people nearby. The company says that if someone steps into a humanoid robot’s workspace, the model stops the machine, then lets it get back to work once that person has moved clear. DeepMind is also introducing a new benchmark meant to test a model’s ability to act as a safety layer over a VLA system: enforcing constraints, monitoring the environment, and asking a human for clarification when a task’s feasibility is unclear.
Gemini Robotics ER 2 is available now through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform. It plugs into the Gemini Live API’s bidirectional streaming endpoint, which DeepMind says removes the “stop and think” pauses that come from calling a reasoning model turn by turn instead of continuously.
The release lands days after AI Insiders reported that DeepMind is reorganizing away from deep-science work like AlphaFold toward the Gemini product line, and robotics is where a visible share of that shift is now landing. For robotics teams, the practical decision is which layer to buy: ER 2 replaces the planning and progress-tracking logic sitting above whatever action model already drives the hardware, not the action model itself. Teams already running a VLA stack should test whether swapping in ER 2 as the orchestration layer cuts retries and task-completion time before committing budget to a broader integration.
Google DeepMind announced Gemini Robotics ER 2 in a company blog post published July 30, 2026.
