Google DeepMind has unveiled Gemini Robotics 2, a suite of AI models aimed at enhancing robot autonomy through whole-body control, advanced dexterity, and multi-robot collaboration. The system is demonstrated on Apptronik's Apollo 2 humanoid, which can now perform full-body movements such as walking, crouching, and bending while reasoning through tasks in real time.
Key takeaways
- Three models: Gemini Robotics 2 (VLA for control), ER 2 (embodied reasoning for planning and collaboration), and On-Device 2 (local VLA for edge deployment).
- Whole-body coordination from 'feet to fingertips' enables tasks like picking and placing objects in specified locations.
- Improved dexterity supports five-fingered hands for tasks like tying bags, plus parallel grippers for industrial insertion and packing.
- Multi-robot collaboration via ER 2 allows different robot types to coordinate on complex workflows.
- On-Device 2 adapts to new dual-arm platforms with fewer than 200 training examples collected in hours.
- ASIMOV-Agentic benchmark evaluates safety, including refusal of unsafe actions and requesting human intervention.
For manufacturers, the suite promises faster adaptation to new hardware and safer human-robot interaction, potentially reducing deployment times for automation tasks.
Source: Robotics & Automation News (roboticsandautomationnews.com) · Published 2026-07-31 · “Google DeepMind unveils Gemini Robotics 2 as Apptronik humanoid demonstrates whole-body AI”