Gemini Robotics 2 Unveiled and Robotics AI Ecosystem Expands

로보틱스 | Fri Jul 31 2026 00:00:00 GMT+0000 (Coordinated Universal Time) | 5 sources

Google DeepMind's whole-body humanoid AI model, surgical robot simulators, and training data collection efforts mark comprehensive progress in physical AI.

Analysis

[Google DeepMind] unveiled Gemini Robotics 2 and announced whole-body humanoid control AI ^[1][2]

• Whole-body motion control from feet to fingertips
• Supports dexterous five-finger manipulation
• Based on vision-language-action (VLA) model
• Supports various humanoids including Apptronik Apollo 2

[Gemini Robotics ER 2] released embodied reasoning model supporting long-horizon task reasoning and multi-robot collaboration ^[2][1]

• Embodied reasoning based on vision-language model (VLM)
• Multi-step task planning lasting several minutes
• Supports collaboration between different types of robots
• Available on Google AI Studio and Gemini Enterprise Agent Platform

[Gemini Robotics On-Device 2] updated on-device VLA model that runs locally without internet ^[2][1]

• Optimized for local execution on robot devices
• Rapidly adapts to entirely new robot embodiments
• Can learn new hardware with just a few hours of data

[Google DeepMind Safety Framework] introduced ASIMOV-Agentic benchmark and multi-layered safety approach ^[1][4]

• Human proximity detection and safe stop functionality
• ASIMOV-Agentic benchmark measures safety of robot control AI collaboration
• Guardrails applied at each model layer
• Positioned as a milestone toward "physical AGI"

[NVIDIA Cosmos-H-Dreams] launched real-time generative simulator for surgical robots ^[3]

• Based on action-conditioned world foundation model
• Runs in real-time on a single RTX PRO 6000 GPU
• Served via FlashDreams streaming inference library
• Specialized for da Vinci Research Kit (dVRK) tabletop suturing

[Shift (Microagi)] collected egocentric data for robot training in exchange for free services ^[5]

• Secures training data by filming first-person hand movements
• Provides free cooking and cleaning services in exchange
• Expanding in New York and San Francisco
• Aims to deploy humanoid robots in real life

Sources

• [1] Google DeepMind’s new AI model can control a robot’s entire body - The Verge AI
• [2] Gemini Robotics 2 brings whole body intelligence to robots - Hacker News
• [3] NVIDIA Cosmos-H-Dreams: Bringing Real-Time Generative Simulation to Surgical Robotics - Hugging Face Blog
• [4] Gemini Robotics 2 Brings Google's AI Into the Physical World - Wired AI
• [5] I Got a Free Meal From a Private Chef—Who Filmed It All to Train Robots - Wired AI