Much of the coverage of humanoid robotics has understandably focused on hardware design. Given the frequency with which their developers toss around the phrase “general purpose humanoids”, more attention should be paid to the first bit. After decades of single-purpose systems, the move to more general systems would be huge. We're not there yet.
The effort to develop a robotic intelligence that can fully exploit the wide range of locomotion opened up by bipedal humanoid designs has been an important topic for researchers. The use of generative AI in robotics has been a hot topic recently. New MIT research indicates how the latter can profoundly influence the former.
The biggest challenge in the way of general purpose systems is training. We have a solid grasp of the best ways to train humans to do different things. Approaches to robotics, while promising, are fragmented. There are many promising approaches, including reinforcement and simulation learning, but future solutions will likely involve combinations of these approaches, augmented by creative AI models.
One of the primary use cases proposed by the MIT team is the ability to gather relevant information from these small, task-specific data sets. The procedure is called Policy Composition (PoCo). Tasks include useful robot actions such as hammering a nail and flipping things with a spatula.
“[Researchers] train a separate diffusion model to learn a strategy, or policy, to accomplish a task using a specific dataset,” notes the school. “They then apply the policies learned by the diffusion models to a general policy. that enable robots to perform multiple tasks in different settings.”
According to MIT, adding diffusion models improved performance by 20 percent. This includes the ability to perform tasks that require multiple tools, as well as learning/acquiring unfamiliar tasks. The system is able to integrate relevant information from different datasets into the sequence of operations required to perform a task.
“One advantage of this approach is that we can combine policies to get the best of both worlds,” says Leroy Wang, lead author of the paper. “For example, a policy trained on real-world data may be able to achieve greater specialization, while a policy trained on simulations may be able to achieve greater generalization.”
The goal of this particular work is the creation of an intelligence system that allows robots to manipulate different tools to perform different tasks. The proliferation of multi-purpose systems will bring the industry one step closer to the general purpose dream.