Safe and Interpretable Autonomous Systems Design: Behavioral Contracts and Semantic-Based Perception

Citation

Cai, Karena Xin (2021) Safe and Interpretable Autonomous Systems Design: Behavioral Contracts and Semantic-Based Perception. Dissertation (Ph.D.), California Institute of Technology. doi:10.7907/w3m8-es32. https://resolver.caltech.edu/CaltechTHESIS:04022021-033321217

Abstract

We are on the verge of experiencing a new, integrated society where autonomous vehicles will become a fabric of our everyday lives. And yet, seamless integration of autonomous vehicles into our society will require vehicles to interface safely with humans in an incredibly complex, fast-paced, and dynamic environment. Premature deployment of these new autonomous systems — without safety guarantees or interpretability of algorithms, could prove catastrophic. How can algorithms governing vehicle behavior be designed in a way that guarantees safety, performance, interpretability and scalability? This is the question this thesis seeks to answer.

First, we present a framework for architecting the decision-making module of autonomous vehicles so that safety and progress of agents can be formally guaranteed. In particular, all agents are defined to act according to what is termed an assume-guarantee contract, which is broadly defined as a set of behavioral preferences. The first version of the assume-guarantee contract is a behavioral profile, which is a set of ordered rules that agents must use to select actions in a way that is interpretable. With all agents operating according to a behavioral profile, the interactions however, are not necessarily coordinated. We then constrain agent behavior with an additional set of interaction rules. The behavioral profile combined with these additional constraints, are what we term a behavioral protocol. With all agents operating according to a local, decentralized behavioral protocol, we can provide formal proofs of the correctness of agent behavior, i.e. all agents will never collide and agents will make it to their respective destinations. Not only does the protocol so defined allow us to make formal guarantees, but it is also designed in a way that scales well in the number of agents and provides interpretability of agent behaviors. Safety and progress guarantees are proven and verified in simulation.

Second, we focus on using information from object classifiers to enhance an autonomous vehicle's ability to localize where it is within its environment. The proposed approach for incorporating this semantic information is based on solving the maximum likelihood problem. With a hierarchical formulation, we are not only able to improve upon the accuracy of traditional localization techniques, but we are also able to improve our confidence in the accuracy of object detection classifications. The improvement in robustness and accuracy of these algorithms are shown in simulation.

Thesis Files