Gemini Robotics 2 gives humanoids whole-body control
Google DeepMind released three physical-AI models that let a single policy drive legs, torso, arms and fingers, and coordinate multiple robots on one task.
Google DeepMind released Gemini Robotics 2 on Thursday, a three-model suite that extends its robotics stack from tabletop manipulation to full-body humanoid control and multi-robot coordination.
The three models
The headline vision-language-action model handles whole-body control, unifying locomotion, torso posture and bimanual manipulation under a single policy. In DeepMind's demonstration, Apptronik's Apollo 2 humanoid receives one instruction — place a watering can into a bin on a lower shelf — then walks to a table, grasps the object, crosses to the shelving unit and completes the placement without the task being decomposed by a human operator. The model also drives the SharpaWave five-fingered hand with 22 degrees of freedom, with DeepMind citing knot-tying and sealing a ziplock bag as dexterity tests.
Gemini Robotics ER 2 is the embodied-reasoning layer, a vision-language model acting as the high-level planner. It handles video understanding with continuous progress tracking, native tool calling including Google Search, and orchestration of the VLA model beneath it. DeepMind reports 57.4 percent accuracy on progress classification across five levels and 91.3 percent on moment-finding with a mean absolute distance of 0.96 seconds, and claims sub-second latency roughly four times faster than competing models. A third model, Gemini Robotics On-Device 2, runs locally on robot hardware and can adapt to an unfamiliar embodiment with a few hours of data.
A single checkpoint controls three embodiments — Apollo 2 with SharpaWave hands, Apollo 2 with Inspire hands, and a Franka arm — and robots hand off subtasks to each other through shared semantic understanding. Partners include Boston Dynamics' Spot and Franka's F3 Duo. ER 2 is available through the Gemini API and Google AI Studio, with the Gemini Enterprise Agent Platform in private preview. DeepMind published a separate safety technical report; the model halts when humans are detected nearby and scores above its predecessor on safety instruction-following and human-proximity benchmarks.
Why it matters
Whole-body control has been the standing gap between language-model-driven robotics demos and machines that operate in unmodified human spaces. Folding locomotion and manipulation into one learned policy — and shipping the reasoning layer as an ordinary API product — moves general-purpose robot control from research artefact toward something third-party developers can build against.