How Google built its Gemini robotics models

Google DeepMind has introduced a new family of Gemini Robotics models designed to enhance the capabilities of robots. These models enable robots to learn and perform complex tasks, such as preparing salads, playing games like Tic-Tac-Toe, and folding origami. The head of robotics, Carolina Parada, highlighted a significant moment when a bi-arm ALOHA robot successfully executed a "slam dunk" with a toy basketball, showcasing the model's ability to understand and perform actions it had never encountered before. The Gemini Robotics models are multimodal, integrating physical actions with outputs like text and audio, allowing robots to adapt to new objects and environments without additional training. The Gemini Robotics-ER model focuses on embodied reasoning, enabling robots to recognize and interact with their surroundings effectively. This approach contrasts with traditional methods that train robots for single tasks, as the Gemini models are trained on a wide range of tasks to promote generalization. The adaptability of these models is crucial for future applications in various industries, including complex environments and human-centric spaces. Google aims to develop robots that can assist with everyday tasks, moving closer to a future where robots are integral to daily life.

- Google DeepMind has launched Gemini Robotics models for advanced robotic capabilities.

- The models enable robots to learn and perform complex tasks without prior exposure.

- Gemini Robotics-ER focuses on embodied reasoning for effective interaction with environments.

- The training approach emphasizes broad task learning over single-task training.

- Future applications include assisting in complex industries and human-centric environments.