Generative AI takes robots one step closer to general purpose

Much of the coverage about humanoid robotics focuses on hardware design. Given the way their developers use the phrase “general-purpose humanoid,” the first part deserves more attention. After decades of single-purpose systems, moving to more generalized systems would be a big leap. We’re not there yet.

Trying to create a robotic intelligence that can fully take advantage of the wide breadth of movements opened up by the bipedal human design has been a hot topic for researchers. The use of generative AI in robotics has been a hot topic lately. New research from MIT shows how the latter can deeply impact the former.

One of the biggest challenges on the path to general-purpose systems is training. We have a solid grasp on the best ways to train humans to perform various tasks. Approaches for robotics, while promising, are fragmented. There are several promising methods, including reinforcement and imitation learning, but future solutions will likely involve combinations of these methods, enhanced by generative AI models.

One of the key use cases suggested by the MIT team is the ability to glean relevant information from these small, task-specific datasets. This method is named Policy Composition (PoCo). The tasks include useful robot actions such as hammering nails and flipping things with a spatula.

,[Researchers] “A separate diffusion model is trained to learn a strategy or policy for completing a task using a specific dataset,” the school explained. “They then combine the policies learned by the diffusion model into a general policy that enables the robot to perform multiple tasks in different settings.”

According to MIT, the inclusion of the diffusion model improved task performance by 20%. This includes the ability to perform tasks that require multiple tools, as well as learn/adapt to unfamiliar tasks. The system is able to combine relevant information from various datasets into a series of actions needed to execute a task.

“One advantage of this approach is that we can combine policies to get the best of both worlds,” says Lirui Wang, lead author of the paper. “For example, a policy trained on real-world data may be able to achieve greater inference, while a policy trained on simulations may be able to achieve greater generalization.”

The goal of this specific task is to create intelligence systems that allow robots to replace different tools to perform different tasks. The proliferation of multipurpose systems will take the industry one step closer to the dream of general purpose.