Researchers at MIT have developed PIGINet, a new system designed to efficiently improve problem-solving abilities of domestic robots, reducing planning time by 50-80%.
Under normal circumstances, domestic robots follow a predefined recipe for performing tasks, which is not always suitable for diverse or changing environments. As MIT describes it, PIGINet is a neural network that takes in "plans, images, goals, and initial facts," and then predicts the probability that the mission plan can be refined to find a feasible motion plan.
The team evaluated the new system's ability to help robots function in the kitchen. They measured the time required to solve the problem with the assistance of PIGINet based on previous methods.
"Because everyone's home is different, robots should be adaptive problem solvers, not just recipe followers. Our key idea is to have a general purpose mission planner generate candidate mission plans, and use deep learning model to select promising mission plans. The result is a more efficient, adaptable, and practical domestic robot that can even flexibly navigate complex and dynamic environments. Moreover, the practical application of PIGINet is not limited to the home, said Zhutian Yang, a Ph.D. student at MIT CSAIL and lead author of the work.
"Our future goal is to further refine PIGINet to propose alternative mission plans after identifying infeasible actions, which will further speed up the generation of feasible mission plans without the need for large datasets to train a general-purpose planner from scratch. We believe that this may Revolutionizing the way robots are trained during development and then in everyone's home."
"This paper addresses a fundamental challenge in implementing a universal robot: how to learn from past speed up the decision-making process in unstructured environments with mobile barriers."