MIT Researchers Simplify Contact-Rich Manipulation Planning with AI
MIT researchers have discovered a new way to simplify contact-rich manipulation planning for robots. By applying an AI technique called smoothing, the researchers can summarize multiple contact events into fewer decisions. This allows a simple algorithm to quickly identify effective manipulation plans for robots.
This breakthrough could potentially lead to the use of smaller mobile robots in factories. These robots would be capable of manipulating objects using their entire arms or bodies, instead of relying on large robotic arms that only use fingertips. This advancement could help reduce energy consumption and lower costs in factories.
Additionally, this technique could be beneficial for robots exploring other planets. By using an onboard computer, these robots could adapt quickly to their environments using contact-rich plans.
The researchers emphasized the importance of leveraging the structure and models of robotic systems in order to accelerate decision-making processes. Rather than relying solely on trial and error, designing physics-based models allows for more efficient planning.
While reinforcement learning has been effective for contact-rich manipulation planning, it requires extensive computation due to the billions of potential contact points that robots must consider. Smoothing techniques, on the other hand, simplify the decision-making process by averaging out intermediate decisions and focusing on core robot-object interactions.
The researchers combined their smoothing model with a fast-search algorithm to improve computational efficiency. The new approach reduced computation time to just one minute on a standard laptop. In simulations and hardware tests, the model-based approach achieved the same performance as reinforcement learning, but in significantly less time.
However, the model developed by the researchers has limitations and cannot handle highly dynamic motions. While it is effective for slower manipulation tasks, it cannot create plans for actions like tossing objects. The researchers plan to enhance their technique to address these challenges in the future.
This research was funded by Amazon, MIT Lincoln Laboratory, the National Science Foundation, and the Ocado Group.