FMitF: Track I: Program Synthesis for Robot Learning from Demonstrations

University Of Texas At Austin, Austin TX

Investigators

Abstract

As robots become more widely available and more capable, end-users of such consumer robots will inevitably expect to be able to teach robots how to perform new tasks. Learning from demonstration (or LfD, for short) is a popular paradigm for this problem, where a user demonstrates how to perform the task, and the robot learns a policy that captures what sequence of actions to perform to complete the task. Most existing LfD techniques rely on neural networks to learn such policies. While promising in some settings, such techniques suffer from key limitations, such as requiring large amounts of training data and lacking interpretability. This project's novelties are in addressing these limitations for robot LfD by combining neural networks (which are very effective for perception tasks) with symbolic learning, which excels at reasoning skills. The project's impacts are 1) introducing a new language to seamlessly merge learning programs consisting of both neural- and symbolic- components, 2) providing guarantees that the learned programs satisfy desired notions of correctness, and 3) allowing such learning to be performed with more realistic, noisy, real-world data. The project's contributions also include training and mentoring of students, developing novel teaching curriculum that integrate robotics with formal methods, and empowering more scalable, safe, and interpretable learning for robots. The research objective of this project is to develop a new LfD paradigm based on program synthesis, with the goal of putting robot learning on a more formal, interpretable, and less data-hungry footing. The key intellectual merit of the project lies in the development of a new set of foundational LfD techniques based on program synthesis. The project will advance the state-of-the-art in robot learning from demonstration by making it possible to learn, in a data-efficient way, programmatic policies that are interpretable and verifiable. The project will also advance the state-of-the-art in program synthesis by developing novel techniques that target the unique challenges of the robotics domain, including noisy and high-dimensional sensor data and uncertain interactions with the environment. In addition, the project will also advance the state-of-the-art in verified learning by considering desired correctness criteria. Finally, the project will make advances in the field of learning from unlabeled demonstrations by learning robot execution policies in the absence of a mapping from states to high-level robot actions. 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.