Researchers at the Massachusetts Institute of Technology, Carnegie Mellon University (CMU), and the University of California, San Diego have created an architecture for a robotic manipulation system that could teach robots to handle pizza dough.
The DiffSkill framework has a "teacher" trajectory optimization algorithm solve each step the robot must follow, then trains a "student" machine learning (ML) neural network that learns abstract concepts for when and how to execute each skill during the task.
The system uses this knowledge to reason out execution strategies for completing the task.
The researchers demonstrated that DiffSkill can perform three complex dough-manipulation tasks in simulations better than reinforcement learning-based ML techniques.
"Our framework provides a novel way for robots to acquire new skills. These skills can then be chained to solve more complex tasks which are beyond the capability of previous robot systems," CMU's Xingyu Lin explained.
From MIT News
View Full Article
Abstracts Copyright © 2022 SmithBucklin, Washington, DC, USA