Key Takeaways
- Google DeepMind introduces Gemini Robotics 2, an advanced AI system.
- The new model can control full-body movements of humanoids and multi-step tasks.
- Robots now perform delicate tasks like tying trash bags with varying success rates.
Google DeepMind has unveiled Gemini Robotics 2, a groundbreaking family of AI models designed to enhance the capabilities of robots. This new system allows for full-body control and improved dexterity in humanoid robots.
Unlike earlier versions that focused on upper-body movements, Gemini Robotics 2 can now walk, crouch, stretch, balance, and manipulate objects based on single instructions. In a demonstration, Apptronik’s Apollo 2 robot successfully placed a watering can into a green bin on the bottom shelf by walking to the table, picking up the can, moving to the shelf, and bending down.
The new system comprises three key models: Gemini Robotics 2, which converts visual and auditory inputs into physical movements; Gemini Robotics ER 2, a higher-level reasoning system that plans multi-step tasks, tracks progress through live video, recovers from failed steps, and coordinates multiple robots working together; and Gemini Robotics On-Device 2, which operates locally without an internet connection.
DeepMind claims that Gemini Robotics 2 can adapt to different robot designs using fewer than 200 examples and a few hours of training data. The system includes the five-fingered SharpaWave hand with 22 degrees of freedom, enabling robots to perform tasks such as tying trash bags, closing ziplock bags, and handling light bulbs.
However, performance remains uneven. According to DeepMind, the success rate for unscrewing a light bulb was 92%, while tying a trash bag succeeded only 40% of the time, and sealing a ziplock bag had a 44% success rate. Screwing in a bulb succeeded 36% of the time.
Google is also working on safety controls for robots through ASIMOV-Agentic, a new benchmark designed to test whether a robot’s reasoning system can reject unsafe actions, recognize when tasks are impossible, and ask humans for assistance when uncertain. Gemini Robotics ER 2 is described as Google's safest robotics model so far.
The launch of Gemini Robotics 2 comes amid growing geopolitical scrutiny of robot hardware. The US recently banned new Chinese-made humanoid and quadruped robots from entering the market due to national security concerns. Google is collaborating with partners such as Apptronik, Boston Dynamics, and Agile Robots for this project.
DeepMind’s Carolina Parada stated that the company aims to build an intelligence layer that can work across different types of robots, paving the way towards what DeepMind calls 'physical AGI' — the long-term goal of creating robots capable of handling a wide range of physical tasks similar to those performed by humans.
