Google DeepMind expands humanoid robotics to full-body coordination

The update allows hardware like Apptronik’s Apollo 2 to master fluid movements from its feet to its fingertips. Beyond physical coordination, the revised Gemini Robotics ER 2 vision-language model improves how robots perceive their environment, enabling them to execute multi-step processes over extended durations. The system now recognizes task boundaries, allowing multiple robots to collaborate on shared objectives, such as clearing a garage or organizing tools.

Safety remains a central pillar of this release. The model incorporates enhanced proximity detection, capable of triggering an immediate stop if a human enters the immediate work zone. Furthermore, Google has refined its on-device capabilities, allowing the software to run locally without an internet connection. This version adapts rapidly to diverse hardware configurations, supporting machines with varying sensors and degrees of freedom to ensure the technology functions across a wider array of robotic platforms.