How Generative AI Brings Robots Closer to Achieving General Purpose Functionality

The Evolution of Humanoid Robotics: From Hardware to Intelligence

Most discussions surrounding humanoid robotics have rightfully concentrated on hardware design. However, the frequent use of the term “general purpose humanoids” by developers highlights the need to focus more on the cognitive aspects as well. Transitioning from decades of single-purpose systems to more generalized solutions represents a significant leap, yet we have not fully arrived at that point.

A central topic among researchers is the quest for robotic intelligence capable of utilizing the extensive range of movements enabled by bipedal humanoid designs. Recently, the intersection of generative AI and robotics has emerged as a hot topic, with new research from MIT suggesting that this combination could significantly transform the field.

One of the primary hurdles in developing general-purpose robotic systems is the challenge of effective training. While we understand the best practices for training humans for various jobs, the strategies for robotics remain somewhat fragmented. Promising methods like reinforcement learning and imitation learning are being explored, and it’s likely that future advancements will involve merging these techniques with generative AI models.

The MIT research introduces a method known as policy composition (PoCo), which compiles relevant information from small, task-specific datasets. This innovative approach includes essential robot actions such as hammering a nail or flipping an item with a spatula.

According to the MIT team, researchers train separate diffusion models to learn specific policies for each task using individual datasets. These policies are then integrated into a comprehensive policy that allows a robot to execute multiple tasks across various environments.

The incorporation of diffusion models reportedly enhances task performance by 20%, enabling robots to not only perform tasks requiring multiple tools but also to adapt to new challenges. This system adeptly links essential information from diverse datasets to create a series of actions necessary for task completion.

Lirui Wang, the lead author of the study, notes, “One of the benefits of this approach is that we can combine policies to get the best of both worlds. For instance, a policy trained on real-world data might exhibit greater dexterity, while one trained in simulation may offer better generalization.”

Ultimately, the aim of this research is to develop intelligent systems that permit robots to switch tools as needed for various tasks. The advancement of multi-purpose systems brings the industry closer to realizing the dream of true general-purpose robotics.