Researchers at the Massachusetts Institute of Technology ( MIT ) have designed a new technique that trains general-purpose robots more efficiently. Inspired by advances in language models such as GPT-4 , the scientists have created Heterogeneous Pretrained Transformers ( HPT ), a system that integrates data from multiple sources and modalities to improve the capabilities of robots.
An AI-based approach that trains general-purpose robots
Robots are trained with specific data and in controlled locations, a time-consuming and resource-intensive process. However, the HPT method developed by MIT changes this process by using a ” core transformer” that unifies visual and proprioceptive data, allowing robots to adapt their behavior to different tasks and environments without starting training from scratch.
According to Lirui Wang, one of the lead authors, the challenge in robotics is not just the lack of data, but the heterogeneity of the data, which has been overcome by the new HPT architecture . This system demonstrated superior performance in simulation and real-world tests, outperforming conventional training by more than 20%. Even on assignments different from the pre-training data, the robots showed improvements in accuracy and adaptability.
To achieve this, the researchers collected a massive dataset that included more than 200,000 robot trajectories , integrating information from visual and proprioceptive sensors. This approach allows robots to perform dexterous movements more effectively, which are crucial for complex tasks.
The MIT team envisions a future where HPT can act as a “ universal robotic brain ,” ready to be downloaded and deployed into different types of robots without the need for additional training. While there is still work to be done, such as improving the model’s ability to process unlabeled data, the potential for this technology is immense.
The development of HPT makes training faster and more cost-effective, while offering adaptability for robots in a world of complex tasks and situations. This project was made possible with support from the Amazon Greater Boston Tech Initiative and the Toyota Research Institute .
Follow us on social media and don’t miss any of our posts!
Loading...
