DeepMind’s embodied AI platform separates perception, reasoning and actuation across distinct computational layers. User instructions are first processed by Gemini Embodied Reasoning, a vision-language model that interprets live video to devise a plan. The resulting plan is handed to the vision-language-action component, Gemini Robotics 2.5, which translates it into concrete motor commands for a robot’s actuators. This architecture enables the system to handle a wide variety of physical tasks.
In laboratory tests the new version outperformed its predecessor by 30 to 100 percent on success metrics and operated 20 to 30 percent faster. Evaluation covered more than two hundred distinct activities, ranging from serving food with tongs to manipulating laboratory pipettes. When paired with Apptronik’s Apollo 2 humanoid, Gemini Robotics 2.5 demonstrated the ability to tie a bow using five-fingered hands, a task traditionally difficult because of the deformable nature of ribbon.
Google extended the evaluation beyond the lab by collaborating with Apptronik and Mercedes-Benz to outfit Apollo 2 robots for automotive part kitting in European factories. In that setting the Gemini-powered system achieved a 96 percent success rate despite the variability of real-world components and human workers. The robots still rely on cloud-based inference, so latency remains a factor even as accuracy and speed improve relative to earlier releases.
A major challenge for embodied AI is the scarcity of physical training data compared with text-based models that ingest the entire web. DeepMind addressed this gap by leveraging teleoperation data and a simulation engine called Gemini 4 Argon. Argon generated thousands of digital-twin variations from real-world camera streams, allowing the model to generalize to new tasks more quickly. Synthetic scenarios were then combined with actual demonstration data to ground the robot’s behavior in reality.
Safety testing for physical AI remains loosely defined, prompting DeepMind to release the open-source ASIMOV-Jailbreak benchmark. The dataset contains 1,500 prompts crafted to coax a robot into unsafe actions, such as attempting to close a bag with a fastener it cannot manipulate. When Argon is included in the inference loop, Gemini Robotics 2.5 attains a 98.8 percent safety score, compared with 73.5 percent for the earlier Gemini Robotics ER 2 model.
Gemini Robotics 2.5 is not yet offered as a general service, but it is deployed with Google’s robotics partners, including Agile Robots, Apptronik and Boston Dynamics. By exposing the system to real-world factories while providing a public safety benchmark, DeepMind aims to steer the broader embodied-AI community toward more reliable and accountable robot behavior.