Gemini Robotics 2 by Google DeepMind can adapt to new robot hardware using fewer than 200 training examples, a drastic cut from the thousands needed by earlier systems. This AI brain doesn’t just control one type of machine but can switch between humanoids and industrial arms, running complex tasks like bipedal walking and multi-robot teamwork.
The system combines three AI models: vision-language-action for movement precision, embodied reasoning for planning multi-step tasks, and an on-device module that lets robots operate independently from cloud servers. Google demonstrated this by running the same AI on Apptronik’s Apollo 2 humanoid and Franka’s robotic arm, proving the system’s hardware-agnostic strength.
By slashing the data needed to train robots and enabling smooth transfer across platforms, Gemini Robotics 2 could shake up the robotics industry. The models are accessible via Google AI Studio, and DeepMind is setting safety benchmarks for autonomous decision-making. This approach might lower entry barriers for robot hardware makers, shifting focus toward software intelligence.
This material is informational and not financial advice.
