Researchers Develop AI That Enables a Robot to Create Complex Plans Using its Entire Hand

For robots, the intricate skill of whole-body manipulation that comes naturally to humans has posed a challenge for machines. But now, MIT researchers, are using the power of artificial intelligence to solve this issue with robotics by introducing a novel approach known as contact-rich manipulation planning.

According to the paper, this technique employs AI-driven smoothing strategies to reduce the complexity of judgments required for effective robotic manipulation planning. With billions of potential contact events to consider, these smoothing methods provide a means to distill the best possible manipulation plan from the vast array of contact occurrences.

Recent advancements in reinforcement learning have helped to push forward new advancements in contact-rich dynamics. This is an achievement that had previously eluded traditional model-based techniques. While the reasons behind reinforcement learning’s success were initially unclear, researchers have been focused on unraveling the underlying factors from a model-based perspective.

A key challenge lies in the intricate nature of contact dynamics, which presents difficulties for planning via touch from a model-based standpoint. So the researchers have proposed a solution, introducing a convex, differentiable model of quasi-dynamic contact dynamics.

This model not only enhances predictive capabilities but also addresses issues related to non-smooth dynamics, rendering traditional linear models inadequate. The integration of contact mode smoothing with sampling-based motion planning marks another significant stride.

While speaking with MIT News, H.J. Terry Suh, an electrical engineering and computer science (EECS) graduate student and co-lead author of a paper on the technique said, “Rather than thinking about this as a black-box system, if we can leverage the structure of these kinds of robotic systems using models, there is an opportunity to accelerate the whole procedure of trying to make these decisions and come up with contact-rich plans.”

The researchers’ approach allows robots to efficiently plan within minutes while maintaining generalizability across diverse environments and tasks—a potent alternative to resource-intensive reinforcement learning methods.

The fusion of mode smoothing and Rapidly-exploring Random Trees, or RRT. This opens the door to efficient global motion planning for high-dimensional, contact-rich systems. This convergence leverages local approximations guided by the Mahalanobis metric, ensuring comprehensive exploration and trajectory optimization.

So what does this mean for those outside of robotics? Well, quite a bit. This new technique promises to help reshape the landscape of robotics by pushing the boundaries of what machines can accomplish in the realm of manipulation and interaction.

In time, it could also lead to advancements in robotics that allow for greater task complexity, reducing the need for physical human labor. Which in theory could allow robots to take on more dangerous tasks while reducing risks associated with human life.