Three models — VLA, ER 2 and On-Device 2 — adding coordinated whole-body control, multi-finger dexterity and teamwork between robots. DeepMind says adapting the models to a new robot body takes hours, not months.
Announced on July 30, Gemini Robotics 2 is a family of three models: VLA (vision-language-action, the one that executes), ER 2 (embodied spatial reasoning, the one that plans) and On-Device 2 (running on the robot itself, offline).
What is new: coordinated control of the whole body — not just the arm — dexterity with multi-fingered hands, and collaboration between several robots working on the same task. DeepMind says adapting the models to a different robotic platform drops from months to a few hours.
Safety: the release introduces ASIMOV-Agentic, a benchmark for measuring safe behaviour in robots acting autonomously.
The partners: Apptronik, Boston Dynamics and Agile Robots — humanoid platforms already in factories and warehouses, not lab prototypes.
Why it matters: "physical AI" moves from promise to calendar. If adapting a model to a new body costs hours, the sector's bottleneck stops being software and goes back to being hardware.
Go to DeepMind's announcement →Source: Google DeepMind
Go to the original source →