RI: Small: Generalizing Learned Manipulation Skills to Unseen Situations by Balancing Uncertainties

点击翻译按钮获取中文摘要

Programming a robot to perform a manipulation task requires a great deal of robotics knowledge and programming skills. Even for an expert, it would take a significant amount of time to carefully craft an algorithm and a controller for a robot to perform a particular task in a specified situation. Giving robots the capability to learn from human demonstrations would remove the need for programming and its cost. However, current approaches can only replicate the human's motion or strategies of the learned task in its demonstrated situations but cannot transfer them to unseen conditions. Since it is unrealistic to teach robots all tasks in every possible situation, robots need to be able to generalize the manipulations kills learned in one situation to unseen situations. The goal of the project is to develop a manipulation skill generalization approach that takes into consideration both the variations in the demonstrations and uncertainties in the unseen situation. By balancing them, the robot can adapt the learned manipulation skills to new situations. The project advances the effort of giving robots the capability to learn and continue their learning in practice. With that capability, robots will be able to perform daily living tasks and provide needed help to people with disabilities and seniors without the cost of programming. The project will also produce new teaching materials for robotics courses and train both graduate and undergraduate students. The main idea of the project is that the robot should learn from the demonstration data in the form of a broader probability representation rather than an optimized but narrowed fitting, so that it is easy to incorporate the probability-based learning from demonstration with the predictions in unseen situations. The project has four objects: learning motion distribution inference model from demonstrations; learning the distributions of two manipulation estimated outcomes and desired outcomes; generating an optimal action by incorporating both motion distributions and outcome distributions; and evaluating the approach on real physical system for daily manipulation tasks. The project not only gives robots a capability to learn and generalize their skills in practice, but also produces new tools for research and applications involving making a decision based on experiences and predictable outcomes.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.