NRI: FND: Physics-based training of robots for manipulation of ropes and clothes

点击翻译按钮获取中文摘要

This project enables collaborative robots to manipulate ropes and clothes in the real world using computer models without constant human monitoring. Seamless integration of robots, as aids to humans, into our daily life and manufacturing environments requires autonomous robotic manipulation of everyday objects. A broad class of these objects have one-dimensional (e.g. ropes) or two-dimensional (e.g. towels) geometry and are highly flexible. This flexibility and deformation can be seen in action everyday while tying shoelaces or folding clothes. Robots must be able to predict this deformation and act accordingly for successful manipulation of such objects. Typically robots are trained to perform these manipulation tasks through repetitive human demonstrations; the robots simply replicate the steps undertaken by the trainer. This project, in contrast, replaces training by demonstration with modeling in computer. The robots will employ numerical simulations to figure out the best policies for manipulation that are robust against uncertainties of the real world, e.g. friction and material defects. As such, a collaborative robot will be able to perform manipulation tasks right out of the box. Areas of application for this framework include typing knots in ropes, securing rigid objects using knots, and folding of clothes.The research objective of this project is to fundamentally understand robotic manipulation of flexible objects (ropes & clothes) using model-based training. The team of researchers will develop physics-based simulation tools for the mechanics of deformable structures and demonstrate the application of fast and efficient simulations to train robots. To overcome the barriers associated with translating models to the real world, the researchers will use simulations, in conjunction with optimization, to formulate policies that are robust against uncertainties, e.g. friction and material defects. The goal is simulation-based training of cobots that is ready for application in the real world; this will largely remove the painstaking training process by physical demonstration required for collaborative robots. An open-source software repository, similar to App Stores for smart-phones, is envisioned that will host training programs for a variety of applications. The strength of this approach will be demonstrated through autonomous folding of towels and tying of knots to secure objects.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.