Particle.news

Google DeepMind Launches Gemini Robotics 2 and ER 2 for Whole‑Body, Multi‑Robot Control

This release adds continuous video reasoning, tool orchestration, on‑device action models, built‑in human‑safety checks to enable diverse robots to plan multi‑step tasks.

Overview

  • Gemini Robotics ER 2, which DeepMind announced on July 30, 2026, is available to developers through the Gemini API and Google AI Studio with a private preview on the Gemini Enterprise Agent Platform.
  • The system separates roles so ER 2 acts as a high‑level embodied‑reasoning 'brain' that watches continuous video, reasons about goals, calls tools and delegates motion to lower‑level vision‑language‑action (VLA) models.
  • DeepMind published performance figures showing 57.4% accuracy for continuous progress classification and 91.3% accuracy with a 0.96 second mean error for moment‑finding, while claiming faster execution than larger model classes.
  • Demos show whole‑body control and finer hand dexterity on partner hardware such as Apptronik’s Apollo 2, Franka platforms and Boston Dynamics’ Spot, and enable multi‑robot handoffs through a shared semantic understanding.
  • DeepMind emphasizes layered safety with a new ASIMOV‑Agentic benchmark and improved human‑proximity stopping, but notes limits remain in task progress accuracy and movement speed and wider real‑world deployment, regulation and generalization will take more work.