Google DeepMind ships Gemini Robotics 2 for whole-body robot control

Google DeepMind released Gemini Robotics 2, an update to the vision-language-action line it introduced in March 2025, extending control beyond arm and gripper manipulation to whole-body movement — walking, crouching and reaching in coordination with hand and finger control. The release comprises three separate models: a vision-language-action model that converts visual and language input directly into motor commands, an embodied-reasoning model for higher-level task planning, and a smaller model designed to run on-device.

DeepMind demonstrated the system on humanoid and non-humanoid hardware, including Apptronik’s Apollo 2 and several tabletop manipulator platforms, on tasks such as multi-step manipulation lasting several minutes and coordination between multiple robots. Reported success rates varied widely by task and difficulty — from the low 30s up to the low 90s in percentage terms across whole-body picking, gripper dexterity and multi-finger manipulation — illustrating that reliable general-purpose manipulation remained unsolved even as planning and perception improved.

Availability was staged: the embodied-reasoning model went into Google AI Studio and a private preview of Google’s Enterprise Agent Platform, while the action and on-device models were limited to early-access partners who applied through a sign-up form. The release continued a pattern, also visible at Boston Dynamics and Figure, of robotics labs pairing large multimodal foundation models with dedicated low-latency control layers rather than relying on a single end-to-end network.