[Zhongtai Machinery] Review of Google's New-Generation Robot Large Model

English summary

Event: Google released the next-generation robot model Gemini Robotics 2. The official website commented that this marks an important milestone on the road to AGI, aiming to solve AI challenges in the physical world. Specifically, there are three models: 1. Gemini Robotics 2: VLA model, converting vision/language into motor control instructions. 2. Gemini Robotics ER 2: reasoning model, responsible for task planning, interaction, and tool invocation. 3. On-Device 2: edge-local inference, adapting to new embodiments with fewer than 200 examples. Characteristics and capabilities: 1. Long-horizon task processing: ER2 tracks task progress based on continuous video streams and adjusts in real time. Long-horizon capabilities are essentially embodied by ER2. 2. Multi-robot collaboration: sharing semantic understanding and communicating and interacting to achieve collaboration. Comments: 1. The three-model technical stack is very similar to the Optimus architecture: grok integration handles task management, FSD is responsible for execution, and AI5-6 chips provide ultra-high-end computing power. ...