Google's DeepMind Robotics researchers are making waves in the AI world with their latest innovation: AutoRT. This game-changing technology is set to redefine how robots understand and respond to human instructions.
Traditionally, robots have been designed to perform one task repeatedly. They're great at what they do, but throw a curveball their way, and they're stumped. AutoRT is about to change all of that.
This cutting-edge system leverages large foundational models to give robots a broader understanding of their environment and tasks. For instance, it uses a Visual Language Model (VLM) to help robots equipped with cameras map out their surroundings and identify objects.
But it doesn't stop there. AutoRT also uses a large language model (LLM) to suggest tasks that the robot can perform, including actions with its end effector. This means robots can understand more natural language commands, reducing the need for hard-coding skills.
The DeepMind team has been busy testing AutoRT over the past seven months, and the results are impressive. The system can manage up to 20 robots at once and a total of 52 devices. They've collected over 77,000 trials, including more than 6,000 tasks.
The team has also unveiled RT-Trajectory, another groundbreaking innovation that uses video input for robotic learning. It overlays a two-dimensional sketch of the robot's arm in action over the video, providing practical visual hints to the model as it learns its robot-control policies.
The training had double the success rate of its RT-2 training, achieving a 63% success rate compared to 29% while testing 41 tasks.
In a nutshell, Google's DeepMind is pushing the boundaries of what robots can do. With AutoRT and RT-Trajectory, robots will not only be able to perform tasks more efficiently but also understand and respond to human instructions more naturally. This is a giant leap forward in the world of robotics, and we can't wait to see what's next!
Made with TRUST_AI - see the Charter: https://www.modelprop.co.uk/trust-ai
Commentaires