Google DeepMind Unveils Gemini Robotics ER 2, an Embodied Reasoning Model for Multi-Robot Collaboration

Google DeepMind has unveiled Gemini Robotics ER 2, the latest version of its embodied reasoning model that acts as a high-level “brain” for robots. Beyond real-time spatial reasoning and multi-step task planning, the new model adds the ability for different types of robots to share semantic understanding and hand off tasks to one another.

Details

• Overview: The latest version of Google DeepMind’s embodied reasoning model, which handles high-level cognition for robots — performing real-time spatial reasoning and task planning to serve as a robot’s “brain” in physical environments
• Video understanding and progress tracking: Watches continuous video feeds to monitor how a task is progressing and adapts if something goes wrong. For complex tasks like tightening a light bulb or tying a trash bag, it can judge whether a step is truly complete before moving to the next one
• Tool orchestration: Coordinates multiple control interfaces — including Vision-Language-Action (VLA) models and navigation APIs — while streaming multimodal data such as video, audio, and text. It integrates with the Gemini Live API for the low latency that physical robot control demands
• Multi-robot collaboration: Different types of robots can now communicate through shared semantic understanding to hand off tasks and complete workflows that no single robot could finish alone. A demo showed Apptronik’s Apollo 2 and Franka’s F3 Duo working together
• Enhanced spatial intelligence: Improvements include better success/failure detection from video feeds, more generalized reading of instruments like digital displays, scales, and thermometers, and stronger visual question-answering
• Performance: Achieves 57.4% accuracy on five-level progress classification, and 91.3% accuracy with a 0.96-second mean absolute distance on moment-finding — running at roughly 4x the speed of competing models. On safety, it outperforms the previous generation (ER 1.6) on safety instruction following and human-proximity detection, successfully halting robots when people get close
• Availability: Available through the Gemini API and Google AI Studio, with a private preview also starting on the Gemini Enterprise Agent Platform
• Announcement: Announced July 30, 2026, by Google DeepMind’s Steven Hansen (Senior Staff Software Engineer) and Peng Xu (Staff Software Engineer)
• Demos: Google also showed demos using Boston Dynamics’ Spot, retrieving objects based purely on natural-language commands

How to try it

• Call Gemini Robotics ER 2 through the Gemini API or Google AI Studio
• Sample code is available on GitHub, which is a good starting point for testing the model’s behavior
• Teams considering enterprise deployment can look into joining the private preview of the Gemini Enterprise Agent Platform