level: technical
Google DeepMind launched Gemini Robotics ER 2, a model that acts as a high-level brain for robots. It enables real-time spatial reasoning, multi-step task planning, and collaboration between different robots. The model watches continuous video feeds to track progress and adapt to errors. It can call tools like Google Search and hands off motor execution to lower-level vision-language-action models. Developers can access it via the Gemini API, Google AI Studio, or the Gemini Enterprise Agent Platform.
Gemini Robotics ER 2 achieves 57.4% accuracy on progress classification, dividing task completion into five levels. For moment-finding, it reaches 91.3% accuracy with a 0.96-second mean absolute distance, operating at four times the execution speed of larger models. It outperforms its predecessor ER 1.6 on tool orchestration across real and simulated robot control. The model also improves safety, halting robots when a person is nearby and resuming only when the area is clear, as shown on new benchmarks.
The model integrates with the Gemini Live API for low-latency bidirectional streaming, avoiding stop-and-think pauses. A demo with Boston Dynamics’ Spot shows natural language commands for fetching objects. Multi-robot collaboration lets diverse machines like wheeled rovers and humanoids share semantic understanding to hand off tasks. Google also introduced a benchmark for safe VLA orchestration, testing constraint enforcement and human clarification. The code and examples are available on GitHub to help developers build physical AI agents.
why it matters: This model gives robots real-time video understanding and multi-step planning, making them safer and more capable in physical environments for developers building AI-powered automation.