Project: #291 Labeling Roads with Different Types of Automated Driving Functional Requirements using Machine Learning Progress Report - Reporting Period Ending: June 30, 2020 Principal Investigator: Ding Zhao Status: Active Start Date: July 1, 2019 End Date: June 30, 2020 Research Type: Advanced Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2019 Mobility21 UTC Progress Report (Last Updated: June 30, 2020, 6:55 p.m.) % Project Completed to Date: 100 % Grant Award Expended: 100 % Match Expended & Document: 100 USDOT Requirements Accomplishments This project aims to label the risk levels of map layouts to help safely deploy automated vehicles (AVs) in various communities regarding road types, geometries, lighting facilities, and human behaviors. The methodology includes using Nonparametric Bayesian methods to identify the driving scenarios as an indicator of the objective risk and leveraging subjective logic to represent the drivers' intention and human subjective risk. The accomplishments during the last two months are 4-fold. First, we identified the dataset containing naturalistic Pittsburgh driving data, the Argoverse Dataset. Secondly, we leveraged the Dirichlet Process Gaussian Process to classify the naturalistic data into different scenarios, the number of which is a sign of the risk level. An alternative indicator of risk level is the number of possible driving primitives corresponding to the map layout. Third, we reviewed subjective opinions which can represent human ambiguity and vagueness towards the current driving situation. Another way to represent human risk sensitivity is the coherent risk measure leaned via inverse Reinforcement Learning. Besides, we reviewed government instructions for deploying AVs and risks defined in the public policy area. This project provided opportunities to dig into theories and algorithms related to Nonparametric Bayesian Methods, Reinforcement Learning and human opinion representation learning. Besides, it provides an interesting and applicable platform for the newly developed algorithms. Impacts By labelling the risk levels of different map layouts, a risk heatmap can be built and is helpful to guide the deployment of autonomous vehicles. Matching different levels of automated vehicles (AVs) with map areas in different risk levels can help improve traffic efficiency, such as reducing the probability that catastrophes happen and avoiding the traffic congestion caused by incapable autonomous vehicles. Currently, one main contribution is that we are the first group to use data-driven methods to define driving scenario risks. Instead of manually designing complex features comprising the overall risk, we build a framework leveraging Dirichlet Process Gaussian Process as an end-to-end way which takes large amounts of naturalistic driving data and outputs the risk level directly. Another main contribution is that besides using objective indicators to define risk, we study the human sensed risk represented by subjective logic, which helps increase the human trust level to AVs. Other We keep updating the traffic-net.org website as a tool to analyze the public data of Pittsburgh. http://traffic-net.org/ We updated the DPGP code for the usage of transportation analysis and modeling: https://github.com/mxu34/DPGP Outcomes New Partners None Issues Automated Vehicles (AVs) are developed by various companies and the core technologies are intellectual properties that are hard to get and test on. Hence instead of directly labelling the functional requirements of AVs, we plan to provide an easy to use simulation platform or detailed experiment instructions with concrete metrics to examine whether the AVs are qualified to be deployed in the real world.