Google DeepMind announced enhancements to its Gemini Robotics AI model on Thursday, expanding its capabilities to control the complete physical systems of humanoid robots. The previous version focused on directing upper-body movements, while Gemini Robotics 2 now supports whole-body coordination spanning from the feet through the fingertips. This advancement enables robots to execute a broader spectrum of physical actions, including walking, crouching, stretching, and object manipulation. Demonstration videos featured Apptronik’s Apollo 2 robot performing tasks such as bending to retrieve a watering can and selecting specific items from shelves.
The model upgrade includes enhanced dexterity through support for more sophisticated five-fingered hands, allowing robots to handle delicate tasks like sealing storage bags, securing trash bags, and removing lightbulbs. Google DeepMind acknowledged that movement speed remains an area for further development but described the update as a meaningful progression toward enabling robots to accomplish intricate, real-world assignments requiring coordinated whole-body movements.
The company simultaneously updated Gemini Robotics ER, its vision-language model responsible for environmental analysis and instruction processing. The second version of Gemini Robotics ER demonstrates improved sustained task execution and enhanced recognition of task initiation and completion points. This iteration also facilitates collaboration between multiple robots of varying designs, as illustrated in footage showing Apollo 2 coordinating with Google’s dual-arm robot during a garage cleaning operation.
Google DeepMind emphasized that Gemini Robotics ER 2 represents the safest version of its robotics model, incorporating improved human detection capabilities, automatic safety protocol activation, and emergency stopping functionality when personnel approach too closely. Additionally, the company refined its Gemini Robotics On-Device Model, enabling faster adaptation to different robot configurations with varying physical designs, sensor arrays, and degrees of freedom, while maintaining the capability to operate independently without internet connectivity.
Article Attribution | Read More at Article Source
Article summary produced by Claude AI