Google Unveils Three Embodied AI Models, Solving Robotic 'Temporal Intelligence'

aigclink · x · 2026-07-31

Google has released three models for embodied AI, aiming to solve 'temporal intelligence'—allowing robots to know not only how to perform a task but also when it is completed.

Gemini Robotics 2: A Vision-Language-Action (VLA) model controlling physical robot movements (closed-source).
Gemini Robotics ER 2: An embodied reasoning model acting as the 'brain,' capable of watching video, planning multi-step tasks, and delegating execution to any VLA (closed-source).
On-Device 2: Runs locally and can adapt to a new robot body within hours (open-source).

Key Breakthroughs in ER 2:

Progress Classification: Categorizes video frames into five progress tiers with 57.4% accuracy.
Precise Timing: Identifies the exact frame for key events (e.g., stopping coffee pouring) with 91.3% accuracy and an average error of just 0.96 seconds, achieving sub-second latency with low compute.
Seamless Interaction: Integrates with the bidirectional Gemini Live API for smooth 'thinking while doing' and natively calls tools like Google Search.
Multi-Robot Collaboration: Demonstrated task handovers between Apollo 2 and Franka F3 Duo, plus voice commands for Boston Dynamics Spot.

ER 2 is now available to developers via the Gemini API and AI Studio.

Original post →

Google Unveils Three Embodied AI Models, Solving Robotic 'Temporal Intelligence'

More from Embodied