MIT researchers have unveiled a groundbreaking method for training general-purpose robots, drawing inspiration from the success of large language models. This new technique promises to make robots more versatile, efficient, and adaptable to a wider range of tasks.
Addressing the Challenges of Robot Training
Training robots to perform multiple tasks in various environments has traditionally been an expensive and time-consuming process. Engineers typically gather task-specific data for each robot, which limits the robot’s ability to adapt to new scenarios. To overcome this limitation, MIT researchers developed a new method that combines diverse data from multiple sources, enabling robots to learn from a wider range of experiences.
This new approach, called Heterogeneous Pretrained Transformers (HPT), aligns data from various domains, including simulations and real robots, into a shared “language” that a generative AI model can process. By leveraging a vast amount of diverse data, this system can train robots to handle a broad array of tasks without starting from scratch every time, resulting in faster and less expensive training processes.
Inspired by Large Language Models
The MIT team drew inspiration from large language models (LLMs) like GPT-4, which are pretrained on massive datasets and later fine-tuned for specific tasks. Similarly, HPT allows robots to generalize knowledge from diverse datasets, making them more adaptable to new environments and tasks. This method improves robot performance by more than 20 percent in both simulation and real-world tests compared to traditional training techniques.
According to Lirui Wang, the lead author and EECS graduate student, the true innovation lies in the ability to unify the various forms of data—such as vision sensors and proprioception (the robot’s sense of its own movement)—into a single input that the transformer model can process. This capability enables the robots to make more dexterous movements and respond to changes in their environment with greater precision.
Real-World Applications and Future Potential
One of the key benefits of HPT is its ability to process data from different types of robots with varying hardware configurations. By building a dataset of over 200,000 robot trajectories across different categories, including human demonstrations and simulations, the researchers successfully trained robots to perform a variety of tasks.
David Held, a robotics expert from Carnegie Mellon University, noted that the technique is a game-changer for scaling robot learning methods, as it allows robots to quickly adapt to new hardware and environments, an essential feature in the rapidly evolving field of robotics.
Toward a Universal Robot Brain
Looking ahead, the researchers aim to further enhance HPT by integrating more diverse datasets and enabling it to process unlabeled data, much like how GPT-4 processes unstructured text. The ultimate goal is to develop a “universal robot brain” that can be downloaded and used for any robot, eliminating the need for extensive retraining.
“This is just the beginning,” said Wang. “We are pushing hard to see how scaling this technique can lead to breakthroughs in robotic policies, similar to the progress we’ve seen with large language models.”
Funded in part by Amazon and Toyota, this research could pave the way for more versatile, general-purpose robots that can seamlessly perform a wide range of tasks in dynamic environments.
Sources: https://news.mit.edu/2024/training-general-purpose-robots-faster-better-1028, https://www.rolandberger.com/en/Insights/Publications/Humanoid-robots-From-science-fiction-to-reality.html
