Google DeepMind announced Gemini Robotics 2 on August 10, introducing an advanced AI intelligence platform that lets robots control their entire bodies with high dexterity and work together on multi-robot tasks [1, 2, 3].

The system consists of three interconnected models: Gemini Robotics 2, a vision-language-action model for motor control; Gemini Robotics ER 2, an embodied reasoning model for communication and planning; and Gemini Robotics On-Device 2, a streamlined local AI designed to run directly on robots without internet connectivity [1, 2, 3].

Gemini Robotics 2 enables humanoid robots to perform natural motions such as walking, crouching, and stretching while manipulating objects with dexterous hands and grippers. The demonstrations feature a five-fingered SharpaWave hand with 22 degrees of freedom, showcasing fine motor skills like tying knots, closing Ziploc bags, emptying dishwashers, picking up watering cans, and screwing in lightbulbs [1, 2, 3].

Unlike teleoperated robots like Elon Musk’s Optimus, DeepMind emphasizes fully autonomous control in real-world environments. The system learned tasks through training on human teleoperation data, video examples, and simulations for each demonstrated skill [3].

Chris Matthieu, VP of developer ecosystem at RealSense, noted the challenge of adapting to dynamic situations: "The hard part isn’t making the first decision — it’s recovering from the hundredth when the world has changed. Doors are closed, objects get moved, people walk into the scene, batteries drain, and sensors become partially occluded" [2].

Combining the vision-language-action and embodied reasoning models allows Gemini Robotics 2 to plan multi-step tasks and coordinate teamwork among multiple robots [1, 2]. Safety is ensured through multi-layered guardrails and a new ASIMOV-Agentic benchmark that detects harmful commands [3].

The platform’s adaptability also lets it integrate with new robotic bodies within hours and operate efficiently offline, enhancing practical deployment [1, 2]. Further demonstrations are available in the release videos published on August 10 showing robots executing complex autonomous tasks [1, 2, 3].