MIT scientists have developed a robot training method, which reduces time and costs, while improving the adaptability of new tasks and environments.
The approach – called heterogeneous Dorthhed (HPT) transformers – combines huge amounts of various data from many sources into a unified system, effectively creating a common language that can process generative AI models. This method means a significant departure from traditional robot training, in which engineers usually collect specific data for individual robots and tasks in controlled environments.
The main researcher Lirui Wang – a graduate of electrical engineering and computer science in MIT – believes that while many cite insufficient training data as a key challenge in robotics, the greater problem lies in a wide range of different domains, modalities and robot equipment. Their work shows how to effectively combine and use all these diverse elements.
The research team has developed architecture that unites different data types, including camera images, language instructions and depth maps. HPT uses a transformer model, similar to powering advanced language models, for visual and proprioceptive input processing.
In practical tests, the system showed unusual results-organizing traditional training methods by over 20 percent in both simulated and real scenarios. This improvement was true, even when the robots encountered tasks differ significantly from their training data.
Scientists have collected an impressive set of data for claim, covering 52 data sets from over 200,000 robot trajectories in four categories. This approach allows robots to learn from many experiences, including interpersonal demonstrations and simulations.
One of the key innovations of the system consists in supporting proprioception (awareness of the robot about its position and movement.
Looking to the future, the team aims to increase the capabilities of HPT to process unknown data, as well as advanced language models. Their final vision is to create a universal brain of a robot that can be downloaded and used for any robot without additional training.
Recognizing that they are at an early stage, the team remains optimist that scaling can lead to groundbreaking development of robotic policies, as well as progress in large language models.
You can find a copy of the article of scientists Here (Pdf)
(Photo Possessed photography)
See also: Jailbreaking ai robots: scientists sound alarm over safety defects
Do you want to learn more about AI and Big Data from industry leaders? Check out AI and Big Data Expo It takes place in Amsterdam, California and London. A comprehensive event is located with other leading events, including Intelligent automation conferenceIN BlockxIN Digital Transformation WeekAND Cyber security and clouds.
Browse other upcoming technological events and internet seminars driven by TechForge Here.
