Google DeepMind has unveiled Gemini Robotics 2, the company’s latest advancement in physical AI and robotics. Released on July 31, 2026, this new model represents a significant leap forward in giving robots intelligent whole-body control, advanced dexterity, and the ability to collaborate with other robots.
Whole-Body Intelligence
Unlike previous robotics models that focused on tabletop manipulation, Gemini Robotics 2 unifies whole-body dynamics under a single vision-language-action (VLA) model. This addresses a persistent limitation in learning-based robotics—the structural decoupling of locomotion and manipulation. Previously, navigation models would steer a robot to a coordinate, then hand control over to a secondary manipulation policy once stationary.
Gemini Robotics 2 pairs deep spatial reasoning with long-horizon planning, enabling robots to map multi-step sequences and complete complex, unfamiliar tasks. The model can control everything from tabletop robots to full-body humanoids.
Three Tiers for Different Needs
DeepMind released Gemini Robotics 2 as three separate models with different access tiers:
- Gemini Robotics 2 (ER 2): The flagship model with the highest reasoning capabilities for general robotics tasks
- Gemini Robotics On-Device 2: A version optimized for running directly on robot hardware
- Safety-focused variant: The team’s safest model yet, with robust ability to understand safety constraints and halt actions when humans are nearby
Multi-Robot Collaboration
A standout feature is multi-robot collaboration. Gemini Robotics 2 allows two robots to divide labor and collaborate on completing a single task—a capability that could prove transformative for warehouse, manufacturing, and logistics operations.
Training with Apptronik’s Robot Park
The model was trained using data from Apptronik’s Robot Park facility, where Apollo 2 robots collect real-world interaction data. This partnership between DeepMind and Apptronik represents a new paradigm for robotics AI training, combining simulation with physical robot data collection.
With Gemini Robotics 2, Google DeepMind is positioning itself as the “intelligence layer” for the next generation of adaptable robots—machines that can move beyond pre-programmed tasks and truly adapt to changing environments.