Particle.news

Google DeepMind Releases Gemini Robotics ER 2 to Developers

The model gives robots continuous video-aware planning, low-latency control, human-proximity safeguards, and a new safety benchmark to guide coordinated multi-robot work.

Overview

  • Google DeepMind published Gemini Robotics ER 2 on July 30, 2026 and made it available to developers through the Gemini API and Google AI Studio with a private preview on the Gemini Enterprise Agent Platform.
  • ER 2 is an embodied reasoning model that watches continuous video, plans multi-step tasks, talks with humans, calls tools, and delegates motor execution to lower-level vision-language-action (VLA) models.
  • DeepMind reports ER 2 scored 57.4% accuracy on progress-classification tests and 91.3% on moment-finding with a 0.96-second mean error, showing big gains in pinpointing task events but remaining limits in progress estimation.
  • The company demonstrated ER 2 coordinating different robots in real hardware demos, including Boston Dynamics’ Spot, Apptronik’s Apollo 2 with SharpaWave hands, and a Franka F3 Duo, to show task handoffs and whole-body dexterity.
  • Google released ASIMOV-Agentic, a benchmark for agentic safety orchestration, and said ER 2 improves instruction following and human-proximity stops; the broader VLA and on-device VLA models are in early-access testing, underscoring a staged rollout and ecosystem approach that could reshape robotic work but will remain partner-driven for now.