Google DeepMind presented Gemini Robotics 2, the new version of its artificial intelligence for physical robots, which promises greater dexterity, real-time analytical capabilities, and more robust safety mechanisms. The launch includes three models, although for now only one is publicly available for developers.
The goal of Gemini Robotics is to create a general-purpose robot, capable of performing any task that a person would do. Google DeepMind scientists refer to this concept as "physical AGI": a system that is instructed on what to do and executes it.
Google DeepMind introduced Gemini Robotics
What is Gemini Robotics ER 2 and what improvements does it bring
Gemini Robotics ER 2 is the updated "embedded reasoning" model, which according to DeepMind represents a significant leap from the previous version 1.6. It is integrated with the Gemini Live API, allowing developers to test its new capabilities.
This is a vision and language model (VLM) designed to understand instructions and the environment. The main improvement is that it can process live video from the robot's cameras, allowing it to track the progress of each task step by step.
Classifies the completeness of video frames with nearly 60% accuracy
Identifies key moments of a task with nearly 90% accuracy
Allows real-time error detection and correction of a single step, without restarting the entire task
For example, if a ball that the robot is trying to grab rolls away, the system can readjust the position and movement of its hand without starting over from scratch.
Robots collaborating with each other
The new embedded reasoning model also gives Google DeepMind's robotics AI the ability to coordinate tasks among different robots. In demonstrations, Apptronik's Apollo 2 and the simpler Franka F3 Duo worked together on the same task without interfering with each other.
Although the test robots still do not match the speed or fluidity of a human, the video demonstrations show behavior with fewer hesitations compared to previous tests.
Robots that collaborate with each other
How the model that executes the robot's actions works
Once the vision and language model plans the task, execution is handled by another system: the updated vision-language-action model, called Gemini Robotics 2. This AI generates the robot's actions based on the instructions, similar to how other generative systems create text or images.
There is also a low-latency version for offline use, called Gemini Robotics On-Device 2. Both models are currently limited to a small group of testers. According to Google, even the on-device version can adapt to new robot designs with just a few hours of movement data, or around 200 examples.
The new safety standards of Gemini Robotics 2
With this version, Google DeepMind incorporated a new safety benchmark called ASIMOV-Agentic, which evaluates models on various safety factors. The test analyzes whether an embedded reasoning agent rejects calls to unsafe tools, whether a task can be completed safely, and whether the model can request human assistance in case of safety doubts.
The new safety standards of Gemini Robotics 2
According to the company, Gemini Robotics ER 2 is the safest model developed to date by the team, with a strong ability to stop actions when a person is too close to the robot. The new safety benchmark is fully available on Hugging Face.
