Generative AI takes robots a step closer to general purpose
MIT researchers advance in creating general-purpose robots using generative AI, improving task performance by 20% with diffusion models.
Generative AI is pushing the development of general-purpose humanoid robots, a significant leap from single-purpose systems. MIT researchers have developed a method called policy composition (PoCo), which combines information from task-specific datasets to train robots effectively in various tasks.
The research shows that generative AI, particularly diffusion models, can improve task performance by 20%. This includes the capability to execute tasks requiring multiple tools and adapt to unfamiliar tasks by integrating data into a coherent sequence of actions.
The ultimate aim of this research is to create intelligent systems that allow robots to switch tools effortlessly for different tasks, moving the industry closer to achieving fully general-purpose robots. By combining policies trained on real-world and simulated data, robots can gain both dexterity and generalization capabilities.