The new architecture manages full-body movements and five-fingered hands. The company reports progress on safety but acknowledges limitations in movement speed.

Google DeepMind has announced Gemini Robotics 2, a new version of its artificial intelligence model designed to control the entire body of a humanoid robot. According to The Verge, the system can coordinate movements from the feet to the fingers, expanding beyond the previous version, which focused primarily on the upper body. This extension broadens the range of tasks robots could potentially perform by combining locomotion, balance and manipulation into a single sequence.

The technical advance lies in the ability to manage actions involving different parts of the robot simultaneously. Gemini Robotics 2 enables humanoids to walk, crouch, reach and handle objects. Operationally, the model must connect body position with arm trajectories and hand grip—a form of coordination required when an object is near the floor, beyond immediate reach or on a shelf.

Google’s demonstrations feature Apollo 2, the humanoid robot developed by Apptronik. In one video, the machine bends down to pick up a watering can; in another, it identifies and retrieves specific items from a shelf, including a baseball glove. The Verge presents these examples based on materials supplied by the company. The available content includes no independent testing, comparisons with competing models or quantitative performance measurements.

Mobility remains an incomplete area, however. Google DeepMind has acknowledged that the robots still need to improve their movement speed, while describing the update as an important step toward more complex real-world tasks. The strategic objective is to move from isolated gestures to assignments in which a machine must relocate, change posture and interact with its surroundings without interrupting the continuity of the action.

Alongside general coordination, the new model also improves dexterity. Gemini Robotics 2 can control articulated five-fingered hands, which are necessary for operations that involve more than simply grasping an object. Examples include sealing a pressure-closure bag, tying a garbage bag and unscrewing a light bulb. Although these activities differ, they all require controlled grip strength and coordinated fine movements.

The announcement also includes Gemini Robotics ER 2, an updated version of the integrated reasoning model that combines visual and language capabilities. Its role is to analyze the robot’s surroundings, process instructions and organize complex, multi-stage tasks. According to Google DeepMind, the new version can sustain activities for longer periods, an important factor in progressing from a single operation to processes made up of consecutive actions.

ER 2 is also said to be capable of recognizing when a task begins and ends. This introduces a level of task management distinct from simple mechanical execution: the robot must interpret the objective, determine which steps remain and decide when the requested outcome has been achieved. The source does not, however, specify the maximum duration of the tests or the success rate recorded for longer sequences.

Another development concerns coordination among robots with different characteristics. According to the company, the update allows multiple machines to collaborate on completing a task. In one demonstration, Apollo 2 instructs a two-armed Google robot to place tools in a container while cleaning a garage. The example shifts the focus from controlling a single humanoid to distributing work among specialized platforms.

On governance, Google DeepMind describes Gemini Robotics ER 2 as its safest robotics model to date. This assessment comes from the company itself. The material reported by The Verge ends before explaining any technical safeguards and does not provide enough information to reconstruct the metrics, protocols or comparative criteria used. There is therefore insufficient evidence to evaluate how safety was tested or which operational risks are covered.

Overall, the announcement presents an expansion of capabilities rather than documented commercial adoption. The source provides no information about customers, costs, availability dates or industrial applications, nor does it offer data that would make it possible to measure reliability outside company demonstrations. Key unanswered questions concern movement speed, continuity during prolonged tasks and the rules governing how multiple robots will coordinate in real-world environments.