Gemini Robotics 2 release notes / what's new

Google DeepMind has released Gemini Robotics 2, an intelligence layer designed to move robots beyond pre-programmed sequences toward adaptable, general-purpose physical AI. The update introduces whole-body control for humanoids, high-precision dexterity for various end effectors, and agentic reasoning for complex, multi-step task execution.

The Gemini Robotics 2 Model Suite

Gemini Robotics 2 is composed of three specialized models that handle different aspects of robotic intelligence, from high-level planning to low-level motor control:

  • Gemini Robotics 2 (VLA): A vision-language-action model that converts visual and linguistic inputs directly into motor control. It supports full humanoids and bi-arm robots, providing control from "feet to fingertips."
  • Gemini Robotics ER 2 (Embodied Reasoning): A vision-language model (VLM) that acts as the agent's high-level brain. It manages human communication, physical world understanding, and the planning of multi-step tasks that can last several minutes.
  • Gemini Robotics On-Device 2 (VLA): An efficient VLA model optimized for local execution to eliminate network latency. It features fast adaptation capabilities, allowing it to be deployed on new robot embodiments with a few hours of data.

Whole-Body Control and Humanoid Motion

Gemini Robotics 2 expands physical AI from upper-body tabletop tasks to full whole-body coordination. This allows humanoid robots to navigate and manipulate objects in environments built for humans, requiring actions such as walking, crouching, and stretching.

In demonstrations using the Apptronik Apollo 2 humanoid, the model successfully processed instructions to move a watering can from a table to a specific bin on a bottom shelf, requiring the robot to walk, pick up the object, and place it precisely.

Advanced Dexterity for Hands and Grippers

The system introduces improved finesse across different types of hardware end effectors:

  • Multi-fingered hands: Using the 22 degree-of-freedom SharpaWave hand on the Apollo 2, the model can perform delicate tasks such as sealing a ziplock bag or tying knots.
  • Parallel grippers: On the Franka Duo platform, the model enables complex dexterous tasks, such as tight packing, using standard two-fingered grippers.

Agentic Reasoning and Multi-Robot Collaboration

Gemini Robotics ER 2 enables robots to execute longer task sequences involving hundreds of decisions over several minutes. The model can now better identify when tasks begin and end and pinpoint key events during execution.

Key capabilities include:

  • Self-Correction: The reasoning model allows robots to generalize to novel situations and self-correct if a specific step in a sequence fails.
  • Multi-Robot Collaboration: Different types of robots can now communicate and coordinate to solve complex workflows that exceed the capabilities of a single robot.

On-Device Adaptation and Motion Transfer

Gemini Robotics On-Device 2 utilizes "motion transfer" techniques to adapt to new robotic hardware quickly. The model can adapt to new bi-arm robot embodiments—even those with significantly different sensors, shapes, and degrees of freedom—using typically fewer than 200 examples and a few hours of adaptation time. This has been demonstrated across the Dexmate, SO101, and Trossen platforms.

Safety Frameworks and ASIMOV-Agentic

Google DeepMind has implemented a multi-layered safety approach combining physical safety measures with AI safety frameworks.

To evaluate these systems, they introduced ASIMOV-Agentic, a benchmark for agentic safety orchestration and uncertainty resolution. This benchmark measures:

  1. The ability of the embodied reasoning agent to refuse unsafe tool calls from a VLA.
  2. The ability to predict if a task is possible and proactively request human intervention when uncertain.

Gemini Robotics ER 2 is reported as the safest robotics model to date regarding human proximity benchmarks and safety constraint following, with the ability to trigger safety tool calls and stop the robot if a human approaches too closely.

Sources