Gemini Robotics 2 Brings Whole Body Intelligence to Robots

Previous models only controlled the upper body of humanoid robots for desktop tasks; Gemini Robotics 2 extends control to the entire body for the first time. On the Apptronik Apollo 2 humanoid robot, the model can follow the command "put the watering can into the green box on the bottom shelf," autonomously performing coordinated actions like walking, grasping, and placing. However, DeepMind acknowledges that the robot's movement speed still needs improvement.

This model is DeepMind's most advanced vision-language-action (VLA) model, converting visual and language inputs directly into motor control signals, capable of controlling full humanoid robots and dual-arm robots. The same model weights can adapt to three different robot morphologies: Apollo 2 with SharpaWave hand or Inspire hand, and Franka Duo with Robotiq gripper.

Gemini Robotics 2 achieves new levels of dexterity on various end-effectors. On the Apollo 2 equipped with a 22-degree-of-freedom five-fingered SharpaWave hand, the model can perform fine actions like tying knots and sealing zipper bags; on the Franka Duo platform with standard two-finger parallel grippers, it can also execute complex tasks such as tight packing.

However, multi-finger dexterous manipulation remains a challenge. Official charts show low success rates for five-fingered tasks, while full-body control and gripper tasks achieve medium to high success rates. DeepMind states it will continue to improve precision and speed to approach human-level dexterity.

Gemini Robotics ER 2 serves as the robot's "advanced brain," responsible for understanding user instructions, observing the environment, planning steps, and coordinating execution with the VLA model. It can execute complex task sequences lasting several minutes and involving hundreds of decisions, and can self-correct and generalize to new scenarios.

This model also introduces multi-robot collaboration for the first time, allowing different types of robots to communicate and cooperate to complete complex workflows that a single robot cannot accomplish. The ER 2 model is now available on Google AI Studio and in private preview on the Gemini Enterprise Agent Platform.

Gemini Robotics On-Device 2 is a VLA model optimized for local operation without network connectivity. It natively supports multiple morphologies, inheriting the "motion transfer" technology from Gemini Robotics 1.5, and can adapt to entirely new dual-arm robot morphologies—including platforms with vastly different shapes, sensors, and degrees of freedom (such as Dexmate, SO101, Trossen)—in just hours with fewer than 200 examples.

DeepMind emphasizes that safety is the cornerstone of robotics research. As physical capabilities increase, they adopt a multi-layered approach combining traditional physical safety measures with robust AI safety frameworks to ensure end-to-end safety and alignment. Specific safety mechanisms are not detailed in this article.