Gemini Robotics 2 extends VLA control from whole-body motion to dexterous hands

Official model update review — source published July 30, 2026

Google DeepMind introduced Gemini Robotics 2 as a family of models for robot control and embodied reasoning. The release separates three functions that are often collapsed into one “robot foundation model” label: a vision-language-action model that outputs robot actions, an embodied-reasoning model that plans and monitors longer tasks, and a smaller VLA intended to run locally on robot hardware.

The important shift is scope. DeepMind reports control across full humanoid bodies, multi-finger hands, parallel grippers, and multiple robots. That does not mean the system has solved general-purpose dexterity. The company’s own task chart shows wide variation, particularly for multi-finger tasks, and the public sources do not establish tactile sensing as an input modality.

What is Gemini Robotics 2?

Gemini Robotics 2 is Google DeepMind’s July 2026 vision-language-action model for turning visual and language inputs into motor actions on humanoid and bi-arm robots. It is presented alongside Gemini Robotics ER 2 for high-level embodied reasoning and Gemini Robotics On-Device 2 for local manipulation inference.

These access states matter. The release is not equivalent to a downloadable, reproducible open model package. Researchers can inspect the announcement, model cards, and safety report, but general access to the VLA and On-Device 2 weights or training data is not described.

What the official task chart reports

DeepMind says one Gemini Robotics 2 checkpoint controlled three embodiments: an Apptronik Apollo 2 with Inspire hands, an Apollo 2 with SharpaWave hands, and a Franka Duo with a Robotiq gripper. The values below are transcribed from the company’s release chart and belong only to those reported task settings.

These figures are useful because they show the difficulty gradient inside “dexterity.” They are not an independent leaderboard. The public announcement does not provide enough information to compare the percentages directly with a different lab’s policy, robot, object set, reset procedure, or success definition. DeepMind itself notes that multi-finger manipulation remains challenging.

Whole-body control and embodied reasoning are different layers

The VLA maps observations and instructions to actions. ER 2 is presented as a higher-level agent that can break a longer goal into steps, communicate with humans, monitor progress, and coordinate tools or other robots. In the release architecture, the reasoning model can call the VLA rather than replacing the controller.

That division is relevant to robot VLA models and Physical AI: long-horizon planning does not remove the need for fast state estimation and low-level control. A robot still needs embodiment-specific balance, collision handling, force control, and hardware safety below the language-reasoning layer.

What the on-device model adds

DeepMind describes On-Device 2 as taking text, images, and numerical proprioception and returning numerical robot actions. The company reports adaptation to new bi-arm embodiments with a few hours of data, typically fewer than 200 examples in its experiments. Its model card also states two important limitations: weaker generalization outside the training distribution and difficulty controlling high-degree-of-freedom robots.

Local inference may reduce dependence on network connectivity and latency, but “on-device” is not itself a latency guarantee. The public model card does not publish a universal end-to-end control frequency across supported robot hardware.

The tactile question: dexterous is not the same as tactile

The release demonstrates hands and grippers, but the public announcement and On-Device 2 model card describe vision, language, and proprioception—not tactile arrays or robot-skin signals—as model inputs. RoboSkin.ai therefore does not classify this release as a tactile foundation model.

That boundary is strategically useful. Multi-finger tasks such as tying, sealing, and tool handling are exactly where contact state, shear, slip, and distributed force can matter. The release establishes a strong vision-to-action reference point; it does not show whether touch improves the reported tasks. A future comparison should document the sensor stack and test the same policy with and without tactile input.

For the contact layer, compare the robot hands guide, tactile AI architecture, and robot manipulation overview.

Safety evidence is also layered

DeepMind’s separate safety report evaluates agentic safety and uncertainty handling, including when an embodied-reasoning agent should refuse, ask for help, or call a safety tool. The report explicitly says it does not evaluate the certified functional-safety architecture, redundancy, or real-time guarantees needed for a compliant physical deployment.

That is the correct system boundary. Semantic safety can decide that a request is unsafe. Hardware and control safety must still make the robot stop predictably when a hazardous condition occurs. See the robot safety map for the distinction between AI behavior, protective sensing, control functions, integration, and validation.

Evidence boundary

This article analyzes Google DeepMind’s own release post, model cards, and safety report. The success rates are provider-reported results, not independent replications, and should not be transferred to other robots or task protocols. RoboSkin.ai did not test the models. No public source reviewed here establishes tactile input, model-weight access for the VLA, general commercial availability, or certified deployment safety.

Sources

• Google DeepMind: Gemini Robotics 2 brings whole body intelligence to robots
• Google DeepMind: Gemini Robotics ER 2 model card
• Google DeepMind: Gemini Robotics On-Device 2 model card
• Google DeepMind: Gemini Robotics 2 safety evaluations