Google launches Gemini Robotics ER 2 for smart robots

Google has released Gemini Robotics ER 2, a new embodied reasoning model designed to serve as what the company calls a "high-level brain" for physical AI agents. Now available via the Gemini API, Google AI Studio, and in private preview on the Gemini Enterprise Agent Platform, this model upgrades the previous Gemini Robotics ER 1.6. It allows robots to process continuous video feeds, plan multi-step tasks, and hand off motor execution to lower-level vision-language-action (VLA) models or navigation APIs.

The model introduces significant performance improvements across several key benchmarks. In progress classification, which tracks task completion across five levels from 0-20% to 80-100%, Gemini Robotics ER 2 achieves 57.4% accuracy. For moment-finding tasks, which identify the exact video frame of a critical event, the model reaches 91.3% accuracy with a 0.96-second mean absolute distance, operating at four times the execution speed of larger model categories. It also outperforms its predecessor on Safety Instruction Following, Human Proximity, ERQA, success detection, and general instrument reading, which was tested across 10 different instrument types.

For robotics practitioners, these upgrades translate to fluid orchestration without jarring pauses. By integrating with the Gemini Live API's bidirectional streaming endpoint, ER 2 enables robots to think while acting. Developers can declare low-level control interfaces as tools, allowing the model to orchestrate complex workflows. Google demonstrated this by pairing ER 2 with Boston Dynamics' Spot to fetch snacks, and by enabling multi-robot collaboration between Apptronik's Apollo 2 and the Franka F3 Duo. The model's enhanced safety features also allow physical agents to autonomously halt operations when humans are nearby and resume once the area is clear.