Updated · The Keyword | Google Product and Technology News · Jul 30
Gemini Robotics Launches ER 2 Model With 91.3% Moment-Finding Accuracy for Multi-Step Robots
Updated
Updated · The Keyword | Google Product and Technology News · Jul 30
Gemini Robotics Launches ER 2 Model With 91.3% Moment-Finding Accuracy for Multi-Step Robots
3 articles · Updated · The Keyword | Google Product and Technology News · Jul 30
Summary
Gemini Robotics made ER 2 publicly available to developers via the Gemini API and Google AI Studio, positioning it as a high-level reasoning layer that plans tasks and hands off execution to lower-level robot control models.
ER 2 is built for real-time physical work: it streams video, audio and text through the Gemini Live API, tracks task progress continuously, self-corrects mid-task and avoids the "stop-and-think" pauses that can disrupt robot actions.
On key benchmarks, the model posted 57.4% accuracy in progress classification and 91.3% accuracy in moment finding with a 0.96-second mean absolute distance, while Google said it runs at 4x the execution speed of larger model classes.
The release also adds multi-robot collaboration, letting different machines share semantic understanding and hand off jobs; Google highlighted demos with Boston Dynamics' Spot and collaboration between Apptronik's Apollo 2 and Franka F3 Duo.
Google said ER 2 also improves safety instruction following and human-proximity performance over ER 1.6, part of its broader push to make embodied AI reliable enough for complex everyday tasks.
Will Google's new AI truly make robots as ubiquitous as smartphones, or will hardware limitations shatter the dream of physical AGI?
Can Google's ASIMOV benchmark actually predict and prevent AI-controlled humanoids from making dangerous, unforeseen mistakes in complex real-world environments?
If Gemini Robotics 2 still requires extensive task-specific training, how close are we really to autonomous robots safely navigating our unpredictable homes?
From Gemini Robotics to Industrial Automation: The Race Toward Human-Level AI and the $370B Robotics Opportunity
Overview
The rapid launch and adoption of Google DeepMind’s Gemini Robotics models marked a turning point, enabling robots to understand and act safely in the physical world. With improved spatial reasoning and new capabilities like instrument reading, these models were quickly integrated into platforms such as Boston Dynamics’ Spot and Atlas, and scaled through major industrial partnerships with Hyundai and Apptronik. Underlying this progress is a dual-system architecture that balances high-level reasoning with real-time control, supported by synthetic data and reinforcement learning for greater reliability. While robots are transforming manufacturing and logistics, high costs and unpredictable home environments mean consumer household robots remain out of reach for now.