Project: #329 Development of Safe, Profitable, and Fair Robotaxi Deployment Strategy Progress Report - Reporting Period Ending: March 31, 2021 Principal Investigator: Ding Zhao Status: Active Start Date: July 1, 2020 End Date: June 30, 2021 Research Type: Advanced Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2020 Mobility21 UTC Progress Report (Last Updated: March 30, 2021, 8:58 p.m.) % Project Completed to Date: 65 % Grant Award Expended: 0 % Match Expended & Document: 0 USDOT Requirements Accomplishments In this phase of the project, we aimed to study the socially compliant driving behaviors based on human driving data. Here we understand social norms as “mental representations of appropriate behavior”. A social norm often consists of social conventions that can be explicitly or implicitly learned under certain social contexts, which serves as guidelines for how humans and autonomous vehicles should behave. The main contributions are listed below: 1. We proposed a method to model interactions between human drivers by formulating the interaction as an inverse general-sum non-cooperative dynamic game. We assumed that the driving trajectories of human drivers interaction constitutes a Nash equilibrium of the general-sum non-cooperative dynamic game. 2. We used nonlinear programming to solve for the human driver’s intrinsic utility function parameters that can best explain the driving trajectories during interaction. The average utility function parameters can be seen as social norms obeyed by human drivers. Due to the difficult nature of the original problem as a coupled bilevel optimization, we replaced the bilevel optimization problem with the necessary conditions of the lower-level optimal solution. 3. We verified our method first on a simple illustrative example and then deployed it on the Next Generation Simulation (NGSIM) real-world dataset. Results showed that we can effectively recover the intrinsic parameters and reconstruct the trajectories in the illustrative example and obtain reasonable trajectory reconstruction on NGSIM. Impacts After processing the human driving data of different driving scenarios such as merging, lane changing, unprotected left turn etc., we can provide important information about what are normal human driving interaction behaviors to the decision-making algorithms of autonomous vehicles. The learned parameters can be used for: 1. Evaluate how far an interaction is from the social norm. 2. Generate AV control input sequence which complies with the social norm. We believe that knowing what kind of driving behavior is socially appropriate can improve the safety and social-compliance level of robotaxis. Other maintain and update the project website http://traffic-net.org Outcomes New Partners n/a Issues n/a