Google DeepMind's Gemini Robotics 2 Adds Whole-Body Control for Humanoid Robots

According to The Verge AI, Google DeepMind announced on July 30 that Gemini Robotics 2 now controls humanoid robot bodies in their entirety—from locomotion to fine-grained hand manipulation. The prior iteration managed only upper-body movements; the new release extends autonomy to walking, crouching, stretching, and five-fingered dexterous tasks. This shift represents a significant step toward robots capable of executing real-world missions that demand coordinated whole-body motion.

Expanded Locomotion and Manipulation Capabilities

The Verge reports that Gemini Robotics 2 enables Apptronik’s Apollo 2 robot to perform demonstrations including bending to retrieve a watering can and retrieving specific objects from shelves. The model’s enhanced five-fingered hand control now supports precision tasks—sealing containers, tying bags, and unscrewing fasteners—that were previously outside the scope of robotic control systems. While Google DeepMind acknowledges movement speed remains an area for advancement, the company frames the release as a pathway toward more ambitious real-world workflows.

Multi-Robot Coordination and Task Understanding

Google DeepMind also upgraded Gemini Robotics ER (Embodied Reasoning) 2, the vision-language component that interprets surroundings and processes instructions. According to The Verge, the new iteration better handles extended-duration tasks and can now disambiguate task boundaries—a technical improvement that enables robots to recognize when one operation concludes and another begins. The model also facilitates heterogeneous robot teams; a demonstration showed an Apollo 2 directing a dual-arm robot to organize tools, illustrating cross-platform task delegation.

Safety and Adaptation

Google DeepMind describes Gemini Robotics ER 2 as the company’s “safest robotics model to date,” with enhanced human proximity detection and automated safety halting when bystanders approach. A third update addresses the Gemini Robotics On-Device Model, which executes inference locally without cloud connectivity. This variant now adapts faster to new robot morphologies—platforms with “drastically different shapes, sensors and degrees of freedom”—reducing the data and tuning burden when deploying to novel embodiments.

Why This Matters

The progression from upper-body control to whole-body autonomy marks a shift in robotics toward generalist behavioral models rather than task-specific pipelines. Teams evaluating humanoid platforms for warehouse logistics, manufacturing, or service roles now have a foundation for multi-step coordination without custom motor control layers. The multi-robot collaboration feature addresses a stubborn adoption barrier: the ability to compose heterogeneous teams without retraining per-robot variants. However, the acknowledged gap in movement speed suggests the models remain optimized for deliberate tasks rather than real-time reactive scenarios—a constraint that will shape near-term deployment in human-shared workspaces.