Google's New Gemini Robotics 2 Platform Allows For 'Intelligent Whole

Key Points:

• Google DeepMind has unveiled Gemini Robotics 2, an AI model capable of controlling humanoid robots to perform tasks such as cleaning, picking up objects, and manipulating items like tape, lightbulbs, and garbage bags with real-time, fully autonomous operation.
• Unlike the original Gemini system that focused on robotic arms and hands, the new version introduces "intelligent whole-body control," enabling more complex and coordinated robot movements.
• The system uses multiple AI models, including a vision language model and two vision language action models, to achieve multimodal understanding and control of the robot's full body and hand movements, though it is trained specifically for the tasks shown.
• Google DeepMind emphasizes safety by implementing guardrails at each AI model layer and introducing the ASIMOV-Agentic benchmark to detect potentially harmful commands, addressing concerns about the risks of autonomous robots interacting in varied environments.
• While the technology marks progress toward physical artificial general intelligence (AGI), Google acknowledges the challenges of AI errors in robotics and is committed to advancing safety measures to mitigate risks associated with autonomous robot operation.