CS Talk - Wenchao Li, Boston University
Title: Towards Assured Autonomy: From System Design to Algorithm
Host: Zhong Shao
Abstract:
The rapid advancement of autonomous systems comes with a growing concern over the safety of these systems when they are deployed in real-life situations. Achieving safety is often a complex task, and requires the system to meet an interlaced set of logical, temporal and real-time constraints. An autonomous vehicle, for example, may employ a machine-learning algorithm for navigating around obstacles, while the controller performing the actual maneuver relies on the underlying middleware providing timely communication. In this talk, I will present some of my research on improving the assurance of autonomous systems from both the system design and the algorithm perspectives.
In the first part of my talk, I will present a systematic framework for designing and verifying multi-rate distributed systems. The framework centers around a declarative language called the Robot Architecture Definition Language (RADL) that is used to capture the system architecture and drive code generation and system integration. I will discuss our experience in applying this framework to the design of high-assurance autonomous vehicles in the DARPA HACMS and BRASS programs. In the second part of the talk, I will present our recent work on safety-aware apprenticeship learning. Apprenticeship learning is an inductive supervised learning methods where the reward function of a Markov Decision Process is unknown to the learning agent and the agent has to derive a good policy by observing an expert’s demonstrations. We propose a novel counterexample-guided approach that can ensure both safety and performance of the learnt policy.