Google DeepMind ships Gemini Robotics 2, taking its robot models to whole-body humanoid control

Google DeepMind has moved its robotics models from tabletop arms to full humanoid bodies. On July 30, 2026 the Alphabet research unit released Gemini Robotics 2, a three-model family that it says can drive a humanoid through walking, crouching, reaching and manipulating within a single policy, and that can be adapted to a new two-armed robot with a few hours of data.

The family has three members. Gemini Robotics 2 is a vision-language-action (VLA) model that turns camera input and natural-language instructions into motor commands for humanoid and bi-arm robots. Gemini Robotics ER 2 is an embodied-reasoning vision-language model meant to sit above the action layer as a planner and coordinator. Gemini Robotics On-Device 2 is a smaller VLA designed to run locally on the robot. According to the model card, On-Device 2 is built on Gemini Robotics 1.5 technology and on-device Gemma models and was trained on Google's Tensor Processing Units.

DeepMind published task-level success rates rather than a single headline score. On Apptronik's Apollo 2 humanoid, whole-body picking succeeded 68.4% of the time from a table, 45.7% from the floor and 76.3% from a shelf. With the 22-degree-of-freedom SharpaWave hand fitted to Apollo, the model unscrewed a light bulb in 92% of trials but screwed one in only 36%, tied a trash bag in 44%, used a dustpan in 32% and handled a ziplock bag in 40%. On a Franka Duo with Robotiq grippers it reached 89.6% on precise insertion, 78.9% on diverse tool kitting and 74.2% on general pick and place.

The company said the VLA can carry out multi-step jobs lasting several minutes and involving hundreds of decisions. It also said the model adapts to new bi-arm embodiments with a few hours of adaptation time, typically with fewer than 200 examples, and named Dexmate, SO101 and Trossen platforms as the test cases. For the on-device variant, the model card reports success on SO101 tasks rising to 53.3% from 6.7% for the first version, and on Dexmate tasks to 75.6% from 33.3%, after extended training.

ER 2 is the part developers can touch today. Google said it is available through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform and code samples on GitHub. The model watches continuous video rather than still frames to judge whether a step has succeeded, scores 57.4% on sorting video frames into five task-progress bands and 91.3% on locating a specific moment in a video, with a mean error of 0.96 seconds. Google said it reads ten types of instruments, from digital displays to liquid thermometers, calls tools such as Google Search natively and streams through the Gemini Live API. The VLA and On-Device models are offered only through an early-access partner program and a trusted tester sign-up.

DeepMind thanked Apptronik, Boston Dynamics and Agile Robots as partners. Google's demonstrations showed ER 2 directing a Boston Dynamics Spot to fetch objects by voice command and coordinating Apollo 2 with a Franka F3 Duo arm in a shared task, which the company described as robots handing off work through a common semantic understanding. Apptronik, founded in 2016, closed a Series A extension of $520 million on February 11, 2026, according to its own announcement.

The release lands in a crowded quarter for robot foundation models. NVIDIA added commercial licensing to its open Isaac GR00T N1.7 humanoid model in April and put a 4-billion-parameter Cosmos 3 Edge world model on Hugging Face on July 20. Skild AI introduced its S1 model on August 18. Physical Intelligence, which raised a Series B in November 2025, has published a series of VLA releases since its first π0 policy in October 2024. DeepMind's distinguishing choice is to split reasoning and action into separate models and to expose only the reasoning layer as a paid cloud product.

Technically, the split mirrors how a human supervisor and a skilled operator divide labour. ER 2 decides what should happen next, checks progress against video and refuses unsafe tool calls. The VLA translates each instruction into joint-level motion at control rates a cloud model cannot reach. DeepMind introduced a new benchmark, ASIMOV-Agentic, to measure whether the reasoning agent declines unsafe tool calls and correctly judges whether a task is possible, and said ER 2 stops a humanoid when a person comes close.

The open questions are reliability and access. Success rates between 32% and 92% on household-style dexterous tasks are research numbers, not factory uptime figures. The On-Device model card itself warns that the model is limited in generalising to out-of-distribution tasks and in controlling high-degree-of-freedom robots, and that mobile and whole-body risks are outside its current scope. Google has not disclosed pricing for ER 2 or a date for general availability of the action models.