NSF-BSF: RI: Small: Learning to plan safely

Washington University, Saint Louis MO

Investigators

Abstract

Robots and autonomous vehicles, in order to achieve the goals they are given, create plans that specify what actions to take. A major impediment to this kind of automated planning is that it requires a description of the environment in which the robot or autonomous vehicle operates, of sufficient fidelity to accurately predict the outcome of those actions. Such accurate descriptions of an environment are difficult to write by hand, and so machine learning techniques have been proposed to automatically construct such descriptions from observations of the environment. These methods to-date have generally not provided any guarantees for the accuracy of the learned description. This leaves open the possibility that the actions of the robot or autonomous vehicle could have unintended consequences. In particular, its actions could violate portions of the specified objectives that were intended to ensure its operation is safe. This project will develop methods for learning descriptions of environments that enable the automated planning methods to guarantee that the resulting plan meets the specified objectives. The project investigates how a planning agent can improve its model of the world by using examples of plan executions to learn the effects of actions and when those actions should be taken. The goal is for the agent to either guarantee that its operation is safe or to detect that the requested operation is impossible to guarantee. The research team will develop algorithms that use the provided observations to build a partial, approximate model of the environment and a meta-model that enables control of the safety and effectiveness of plans produced using the models. The team will then develop automated planning algorithms that use these learned models to produce plans with the desired guarantees. The project combines model-based and data-driven methods to generate "safe" plans. When safe plans cannot be guaranteed, the project aims to quantify the probability that a generated plan is "approximately safe" using concepts from probably approximately correct (PAC) learning theory. The project will establish when it is possible to guarantee that plans will succeed, and will furthermore determine the effect of tolerating a small probability of failure in formulating plans for a broader range of circumstances. This research is expected to increase range of domains in which automated planning can be applied, including more situations in which world models are difficult to obtain and safety is a requirement. 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.