Google DeepMind released Gemini Robotics ER 2, a new embodied-reasoning model built to give robots a real-time decision layer for multi-step physical tasks. Rather than controlling a robot's motors directly, the model reads what it sees on a live video feed, breaks a spoken instruction into a sequence of steps, and hands each step to a separate vision-language-action model running on the robot. It can also call outside tools, including Google Search, while a task is still underway.
The design targets a specific gap in practical robotics: models that reason well but move too slowly or too rigidly to work alongside people in real environments. Google DeepMind built Gemini Robotics ER 2 to run inside the Gemini Live API's bidirectional streaming connection, letting a robot keep planning its next move while still carrying out the current one instead of pausing between actions.
That continuous-video approach is also what lets the model track a task's progress instead of simply checking whether it finished. According to Google DeepMind, Gemini Robotics ER 2 sorts each video frame into one of five completion bands and reached 57.4% accuracy on that progress-tracking benchmark, ahead of the earlier Gemini Robotics ER 1.6 model and the rival systems the company tested against. On a separate benchmark measuring how precisely a model can spot the exact frame where a step finishes, such as the moment a cup stops filling, Google DeepMind reported 91.3% accuracy with an average timing error of 0.96 seconds, a result it says matches larger, slower models while running about four times faster.
The release also adds multi-robot collaboration, letting separate machines coordinate on tasks that need more than one body to complete, such as a wheeled rover working indoors alongside a humanoid built for rougher terrain. Google DeepMind also built a demonstration with Boston Dynamics in which Gemini Robotics ER 2 directed a Spot robot's navigation and manipulator arm to retrieve a snack on a spoken command, with the underlying code posted publicly on GitHub.
Gemini Robotics ER 2 is available now through the Gemini API and Google AI Studio, with a private preview running on the Gemini Enterprise Agent Platform for developers building physical AI agents.
