Gemini Robotics 2 brings whole body intelligence to robots — Google DeepMind

Google DeepMind has introduced Gemini Robotics 2, a new AI model designed to provide robots with intelligent whole-body control, fine dexterity, and collaborative capabilities. Unlike traditional robots limited to pre-programmed tasks, this model enables robots to adapt to unpredictable environments and transfer learned skills across different robotic bodies. The system aims to address longstanding challenges in robotics by allowing robots to reason through movements, such as walking, crouching, and manipulating objects in cluttered spaces.

Gemini Robotics 2 demonstrates its capabilities through various robotic embodiments, including the Apptronik Apollo 2 with SharpaWave hands and Franka Duo with Robotiq gripper. The model achieves medium to high success rates in whole-body and gripper-based tasks, though multi-finger dexterous manipulation remains a challenge. The reasoning model, Gemini Robotics ER 2, is now available on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, with early-access partner models including VLA and On-Device versions.

The update expands physical AI into whole-body motions, enabling robots to perform complex tasks like placing a watering can on a shelf. It also enhances dexterity, allowing robots to use multi-fingered hands for delicate actions such as tying knots or sealing bags, and standard grippers for tasks like tight packing. The model processes user instructions, coordinates actions, and tracks progress, enabling robots to execute multi-step tasks lasting several minutes with hundreds of decisions while self-correcting when necessary.

Safety remains a priority, with Gemini Robotics 2 introducing ASIMOV-Agentic, a benchmark for agentic safety and uncertainty resolution. The model improves safety by detecting nearby humans, triggering safe stops, and refusing unsafe tool calls. It also supports multi-robot collaboration, allowing different robots to work together on complex workflows. The system is designed to operate locally on-device, adapting to new robotic embodiments in hours with minimal examples, marking a step toward general-purpose robotic intelligence.