Most reports on humanoid robots have understandably focused on the hardware design. Given how often their developers throw across the phrase “general-purpose humanoids,” more attention needs to be paid to the primary part. After a long time of single-purpose systems, the jump to more general systems might be a giant one. We're just not there yet.
The pursuit of robot intelligence that may fully exploit the big selection of movements enabled by the bipedal humanoid design is a key issue for researchers. The use of generative AI in robotics has also been a hot topic recently. New research from MIT suggests that the latter could profoundly impact the previous.
One of the largest challenges on the road to universal systems is training. We have a solid understanding of the most effective methods to coach humans for various tasks. Robotics approaches, while promising, are fragmented. There are many promising methods, including reinforcement and imitation learning, but future solutions will likely involve mixtures of those methods, complemented by generative AI models.
One of the important thing use cases proposed by the MIT team is the flexibility to collect relevant information from these small, task-specific data sets. The method is known as policy composition (PoCo). The tasks include useful robot actions corresponding to hammering a nail or flipping things over with a spatula.
“(Researchers) train a separate diffusion model to learn a technique or policy for completing a task given a particular dataset,” the varsity notes. “They then mix the policies learned by the diffusion models right into a general policy that permits a robot to perform multiple tasks in several environments.”
According to MIT, incorporating diffusion models improved task performance by 20%. This includes the flexibility to perform tasks that require multiple tools in addition to learning/adapting to unfamiliar tasks. The system is capable of mix relevant information from different data sets into a sequence of actions required to finish a task.
“One of some great benefits of this approach is that we are able to mix strategies to get the most effective of each worlds,” says the paper's lead writer, Lirui Wang. “For example, a technique trained on real data could achieve greater dexterity, while a technique trained on simulations could potentially achieve greater generalization.”
The goal of this specific work is to develop intelligent systems that allow robots to swap different tools to perform different tasks. The proliferation of multi-purpose systems would bring the industry one step closer to the dream of general-purpose use.