Gemini Robotics ER 2 Enables Simultaneous Thinking and Collaboration in Robots

Summary

Google DeepMind launched Gemini Robotics ER 2 on July 30, 2026. This high-level reasoning model enables robots to plan the next step while executing the current one, coordinate across heterogeneous robot types, and invoke external tools in real time, all with sub-second latency. The model is currently accessible via the Gemini API, AI Studio, and a private preview enterprise platform.

Why it matters

The model directly addresses two long-standing bottlenecks in embodied AI: knowing when a physical task is truly complete and enabling multiple robots to share a semantic understanding of a common goal without a central orchestrator. By separating ER 2's 'thinking' function from the 'action' function in low-level VLA controllers, the model becomes robot-agnostic. This could accelerate deployment across a wide range of platforms—from quadrupeds to arms and humanoids—rather than requiring a custom model for each.

Highlights

• 4x faster than similar large models, with 57.4% task classification accuracy and 91.3% event detection success (0.96s average deviation)
• Multi-robot demo: Apptronik Apollo 2 humanoid + Franka F3 Duo arm completing a task neither could do alone by sharing semantic task state
• Architecture separates high-level planning (ER 2) from motor control (VLA models), making the 'brain' portable across Spot, Apollo, Franka, or any VLA-compatible robot

Critical take

The task completion classification score is 57.4%. Although presented as 'ahead of all competitors,' this means the model misjudges task completion approximately 43% of the time, representing a significant reliability gap for safety-critical industrial or hospital environments. The article acknowledges the gap between controlled demos (e.g., Spot fetching a snack) and uncontrolled real-world conditions (variable lighting, unpredictable humans, uneven surfaces). Additionally, since pricing or SLA details have not been published, the commercial viability of sub-second video stream inference remains largely unverified.