Gemini Robotics

Gemini Robotics, launched by Google DeepMind on March 12, 2025, extends Gemini 2.0 into the physical world by adding actions as a new output modality. The family pairs a vision-language-action (VLA) model that converts visual input and natural-language instructions into motor commands with Gemini Robotics-ER, an embodied-reasoning variant that plans multi-step tasks, calls tools, and produces human-readable reasoning chains before acting.

The lineup grew fast: an On-Device variant for local low-latency inference (June 2025), 1.5 with cross-embodiment generalization (September 2025), and a version 2 with full-body humanoid control. Trusted testers include Boston Dynamics, Agility Robotics, and Apptronik. Access remains gated: the ER models reached developers via the Gemini API, while the full VLA stack runs through an early-access program.