On July 31, 2026 (Beijing time), Google DeepMind released a series of three models under Gemini Robotics 2, expanding the control range for humanoid robots from upper body movements in previous generations to coordinated movements involving feet, torso, arms, and fingers. The official announcement was made at midnight Beijing time, with The Verge's report summarizing this change at 1:18:45 AM. Officially, the update is divided into three levels: motion control, embodied reasoning, and device-side operation, currently available in public tools, private preview, and early access.
Three Models and Full-Body Movements
Gemini Robotics 2 within the series is a visual-language-action model responsible for converting camera-viewed environments and language commands into motion control; Gemini Robotics ER 2 is an embodied reasoning model tasked with understanding the environment, communicating with humans, and planning multi-step tasks. Gemini Robotics On-Device 2 is designed for local operation on robot devices and rapid adaptation after changing robot bodies. These positions are based on official Google DeepMind statements, not third-party test conclusions.
In the publicly released Apollo 2 demonstration, upon receiving instructions to place a watering can into the bottom green box, the system directed the robot to walk towards the table, pick up the object, move to the shelf and complete placement. The company uses this example to illustrate that models can convert a single intent into continuous actions involving walking, stretching, grasping, and balancing. However, Google also acknowledges that the robot's movement speed needs improvement; this experiment does not provide deployment results in factories, homes, or customer sites, nor can it determine long-term operational reliability.
Dexterous Manipulation and Long-Task Reasoning
Hand and gripper control is another set of official demonstrations. Google claims that the model controls the five-fingered SharpaWave hand on Apollo 2, which has 22 degrees of freedom, with tasks including tying knots and closing self-sealing bags; on the Franka Duo platform, the model also controls a two-finger parallel gripper to complete tight packing. Charts illustrate average success rates for full-body movements and gripper skills across multiple similar tasks, with multi-finger tasks listed separately. The company summarizes performance in these categories as moderate to high but explicitly states that dexterous multi-finger manipulation remains challenging without disclosing human operation benchmarks for direct comparison.
Gemini Robotics ER 2 handles longer tasks by reading user instructions and observing the room, determining steps, calling on visual-language-action models for execution, and continuously tracking task progress. Google claims that if a step fails, the system can re-plan, determine when a task starts or ends, and locate key event moments. The official work span is several minutes involving hundreds of decisions, but the announcement does not list specific success rates, failure distributions, or independent replication experiments for these long-sequence tasks.
Collaboration Demonstrations and Device-Side Adaptation
Google also includes multi-robot collaboration in ER 2's responsibilities, allowing different types of robots to communicate and share processes that a single device cannot complete independently. Another demonstration involves the same Gemini Robotics 2 model checkpoint controlling three configurations: Apollo 2 with SharpaWave hand, Apollo 2 with Inspire hand, and Franka Duo with Robotiq gripper. This confirms experimental configurations and official capability descriptions but does not represent multi-robot systems entering commercial projects or production environments.
The device-side version targets scenarios with network latency or no internet connection for robots. Google claims it inherits motion transfer technology from Gemini Robotics 1.5 and natively supports different bodies. The company's adaptation claim is that data collected in a few hours, typically less than 200 examples, can adapt to new dual-arm robots with significantly different shapes, sensors, and degrees of freedom. This figure is self-reported by Google without third-party verification; the announcement also does not disclose model size, device computing power, energy consumption, adaptation costs, or supported hardware lists.
Open Range and Safety Boundaries
The three models are not open in the same way. Gemini Robotics ER 2 is already available on Google AI Studio and offers private previews through the Gemini Enterprise Agent Platform; the motion control-focused Gemini Robotics 2 and device-side versions are only provided to early access partners. The official announcement does not provide a comprehensive opening date, pricing, call quotas, commercial terms, or regional scope, so this research release cannot be written as generally available services.
The safety section introduces a test benchmark named ASIMOV-Agentic to measure embodied reasoning agents' ability to refuse unsafe tool calls, judge task feasibility, and proactively request human intervention when uncertain. Google claims the overall solution combines traditional physical protection measures with an AI safety framework, stating that ER 2 can detect nearby personnel, trigger safety tools, and stop robots when proximity is too close but does not provide independent certification or public production records. The next verifiable points are broader opening arrangements, third-party benchmarks, real deployment data, and whether speed and multi-finger operation limitations have improved.