Google DeepMind has introduced Gemini Robotics 2, a new family of AI models designed to give robots full body control, improved dexterity, and the ability to work together. Unlike earlier versions that focused mainly on upper body movement, Gemini Robotics 2 can control an entire humanoid, with Google saying robots can now walk, crouch, stretch, balance, and manipulate objects from a single instruction. In one demonstration, Apptronik’s Apollo 2 was told to place a watering can in a green bin on a bottom shelf, and the robot walked to the table, picked up the can, moved to the shelf, and bent down to place it in the requested location.
The new system includes three models working together. Gemini Robotics 2 functions as the vision-language-action model that converts what a robot sees and hears into physical movements. Gemini Robotics ER 2 serves as the higher level reasoning system, capable of planning multi-step tasks lasting several minutes, tracking progress through live video, recovering from failed steps, and coordinating multiple robots working together, while also being able to use tools such as Google Search. This model is available through the Gemini API and Google AI Studio, though its Enterprise Agent Platform access remains in private preview. Gemini Robotics On-Device 2 runs locally without requiring an internet connection, and DeepMind says it can adapt to a substantially different robot design using fewer than 200 examples and a few hours of training data.
Gemini Robotics 2 can control the five fingered SharpaWave hand, which has 22 degrees of freedom, with Google demonstrating tasks including tying a trash bag, closing a ziplock bag, and handling light bulbs. Performance, however, remains uneven. DeepMind reported a 92 percent success rate for unscrewing a light bulb, compared with 44 percent for tying a trash bag and 40 percent for sealing a ziplock bag, while screwing in a bulb succeeded only 36 percent of the time. Google acknowledged that movement speed and multi-finger dexterity still need significant improvement. The release has been described as another step toward what DeepMind calls physical AGI, the long term goal of creating robots capable of handling the wide range of physical tasks humans can perform, with DeepMind robotics head Carolina Parada saying the company wants to build an intelligence layer that can eventually work across different types of robots.
DeepMind also introduced ASIMOV-Agentic, a new benchmark designed to test whether a robot’s reasoning system can reject unsafe actions, recognize when a task is impossible, and ask a human for assistance when uncertain. Google says Gemini Robotics ER 2 is its safest robotics model so far, capable of detecting when a person moves too close, stopping the robot, and resuming work once the area is clear. The launch comes as robot hardware itself faces growing geopolitical scrutiny, with the United States recently moving to block new Chinese-made humanoid and quadruped robot models from entering the market over national security concerns. Google is working with partners including Apptronik, Boston Dynamics, and Agile Robots, though for now Gemini Robotics 2 and On-Device 2 remain limited to early access partners, while ER 2 is available to developers.
Follow the SPIN IDG WhatsApp Channel for updates across the Smart Pakistan Insights Network covering all of Pakistan’s technology ecosystem.
