Latest30 Jul 2026Robot model (AI)Preview
Gemini Robotics 2
Google DeepMind's most advanced vision-language-action model: it turns camera input and spoken or written instructions into motor control, and can drive full humanoids from feet to fingertips as well as two-armed robots. Demonstrated on Apptronik's Apollo 2 humanoid with Sharpa and Inspire hands and on the Franka Duo. Available to early-access partners.
- Whole-body control of humanoids, plus bi-arm robots
- Published results on Apollo 2: 76.3% pick from shelf, 68.4% pick from table, 92% unscrew bulb
- Access: early-access partners; trusted tester sign-up
30 Jul 2026Robot model (AI)Preview
Gemini Robotics ER 2
Embodied-reasoning model that acts as a robot's high-level planner: it understands scenes and video, talks with people, plans multi-step tasks lasting several minutes and hands steps to a lower-level action model. Publicly available to developers in preview through the Gemini API and Google AI Studio; demos used Boston Dynamics' Spot, Apptronik's Apollo 2 and a Franka Duo.
- Model ID: gemini-robotics-er-2-preview
- Available via Gemini API and Google AI Studio; private preview on Gemini Enterprise Agent Platform
- 57.4% accuracy on progress classification; 91.3% on moment-finding
30 Jul 2026Robot model (AI)Preview
Gemini Robotics On-Device 2
DeepMind's most efficient vision-language-action model, built to run locally on the robot rather than in the cloud. It takes text, images and the robot's own joint readings and outputs actions, and adapts to new two-armed robots in a few hours, typically with fewer than 200 examples. Available to a select group of trusted testers; evaluated on SO101 and Dexmate robots.
- Runs locally on robotic devices
- Adapts to new bi-arm robots with typically less than 200 examples
- Mobile platforms and whole-body control outside current scope
- Access: trusted testers only
14 Apr 2026Robot model (AI)Preview
Gemini Robotics-ER 1.6
Upgrade to DeepMind's embodied-reasoning model for robot planning and perception: better pointing, counting and success detection across camera views, plus a new skill for reading gauges and sight glasses, developed with Boston Dynamics' Spot robot. Offered to developers in preview through the Gemini API and Google AI Studio; succeeded by ER 2 in July 2026.
- Model ID: gemini-robotics-er-1.6-preview
- Instrument reading: 86% success, 93% with agentic vision (ER 1.5: 23%)
- Available via Gemini API and Google AI Studio