The MIT team has suggested a method called Policy Composition (PoCo) that can collate relevant information from small, task-specific datasets. This method has improved task performance by 20%, including the ability to execute tasks requiring multiple tools and learning/adapting to unfamiliar tasks. The ultimate goal of this research is to create intelligent systems that allow robots to use different tools to perform various tasks, bringing the industry closer to the dream of general-purpose systems.
Key takeaways:
- The use of generative AI in robotics is a hot topic, with new research from MIT suggesting it could greatly impact the development of general purpose humanoid robots.
- One of the main challenges in creating general purpose systems is training, with current approaches being fragmented and likely to require combinations of methods, including reinforcement and imitation learning, enhanced by generative AI models.
- The MIT team has developed a method called Policy Composition (PoCo), which allows a robot to perform multiple tasks in various settings by combining policies learned from different datasets.
- The goal of this research is to create intelligence systems that allow robots to swap different tools to perform different tasks, bringing the industry closer to the dream of general purpose systems.