MIT Researchers Develop Policy Composition Technique to Train Multipurpose Robots
MIT researchers have developed a groundbreaking technique to train robots to use tools and perform repairs around the house with unprecedented efficiency. The method, known as Policy Composition (PoCo), combines multiple sources of data across domains, modalities, and tasks using generative AI models called diffusion models.
Traditionally, robots have been trained using small, task-specific datasets, limiting their ability to adapt to new tasks in unfamiliar environments. However, with PoCo, robots can now learn a general policy that enables them to perform multiple tasks in various settings. This innovative approach led to a 20 percent improvement in task performance compared to baseline techniques.
The researchers trained separate diffusion models to learn strategies for completing specific tasks using different datasets. They then combined these policies into a general policy that allows robots to perform a wide range of tool-use tasks. This method not only improved task performance but also enabled robots to adapt to new tasks they had not seen during training.
The team tested PoCo in simulations and real-world experiments, where robots successfully performed tasks such as using a hammer to pound a nail and flipping an object with a spatula. The results were impressive, with the composed trajectories outperforming individual policies.
In the future, the researchers plan to apply this technique to long-horizon tasks that involve using multiple tools sequentially. They also aim to incorporate larger robotics datasets to further enhance performance.
Overall, PoCo represents a significant advancement in robotics training, offering a promising solution to the challenge of combining heterogeneous datasets effectively. With the support of funding from organizations like Amazon, the U.S. National Science Foundation, and the Toyota Research Institute, this research paves the way for the development of more versatile and adaptable robots in the future.
