Details
- Google DeepMind introduced Gemini Robotics 2, an AI "intelligence layer" that advances robots from tabletop tasks to whole-body control, fine dexterity, and multi-robot collaboration.
- The release comprises three models: Gemini Robotics 2 (vision-language-action), Gemini Robotics ER 2 (embodied reasoning vision-language model), and Gemini Robotics On-Device 2 (efficient VLA for local execution).
- Gemini Robotics 2 can control full humanoids and bi-arm robots from feet to fingertips, performing tasks like walking, crouching, stretching, and manipulating objects with both five-finger hands and parallel grippers.
- Gemini Robotics ER 2 acts as a high-level agentic brain, handling natural language instructions, multi-minute multi-step tasks, progress tracking, self-correction, and coordinating different robots to work as a team.
- Gemini Robotics On-Device 2 is optimized for low-latency, offline use and can adapt to entirely new robot embodiments in a few hours using fewer than 200 examples, while the ER 2 model is available via Google AI Studio and Gemini Enterprise Agent Platform.
Impact
By combining whole-body control, advanced dexterity, agentic planning, and fast on-device adaptation, Gemini Robotics 2 pushes physical AI beyond narrow, pre-programmed workflows toward general-purpose robotics. Over the next 12–24 months, this architecture could accelerate deployment of humanoids and collaborative robots in logistics, manufacturing, and service environments, while its ASIMOV-Agentic safety benchmark and proximity-aware behaviors shape emerging standards for human-robot collaboration.
