Project: #365 Towards a Smart, Safe, and Sustainable Sidewalk: A Quantitative Analysis on How Sidewalk Infrastructure Affect Personal Delivery Devices Progress Report - Reporting Period Ending: March 30, 2022 Principal Investigator: Ding Zhao Status: Active Start Date: July 1, 2021 End Date: June 30, 2022 Research Type: Advanced Grant Type: Research Grant Program: FAST Act - Mobility National (2016 - 2022) Grant Cycle: 2021 Mobility UTC Progress Report (Last Updated: March 30, 2022, 7:23 p.m.) % Project Completed to Date: 60 % Grant Award Expended: 0 % Match Expended & Document: 0 USDOT Requirements Accomplishments The task 1’s objective is to build the delivery robot platform with advanced 3D sensors and computational resources. The task 2’s objective is to enable the robot’s capability to collect sidewalks' data efficiently. The task 3’s objective is building a digital-twin simulation environment with the collected data to facilitate following research. We have finished an alpha version of the physical robot platform development, which includes 3 DJI Livox solid-state LiDARs, 1 2D mechanical LiDAR, 1 stereo depth camera, 1 robot chassis with wheel odometer, and 1 Nvidia Xavier computer. We have also developed a general autonomy stack for increasing the delivery robot data collection efficiency, including the sensing data preprocessing module, localization module, perception module, planning and control module. In Particular, we have finished the calibration of all these sensors, which can provide accurate 3D measurement of the surrounding environment of the robot. We further implemented a 3D mapping module based on the Simultaneous Localization and Mapping (SLAM) technology, which can scan the sidewalks accurately and represent the environment as a dense 3D point cloud for future sidewalk infrastructure quality measurement. We have also developed a perception module, which can detect the dynamic obstacles, such as pedestrians, around the robot. The perception results can help to 1) enhance the mapping quality by filtering out dynamic obstacles and 2) record pedestrian trajectories so that we could investigate the influence of different delivery robot autonomous navigation polices on pedestrians. We have implemented a three-level planning and control module. The first level is a global router, which can plan a collision free path based on static map and A* algorithm. The second level is a local planner that can generate smooth and safe trajectories based on the waypoints generated from the router, and we have implemented a sampling-based local planner - DWA, and two optimization-based planners - iLQR and MPC. The third level is the controller that can convert the trajectory to motor commands of the robot. With the three-level planning and control module, we can enable the robot to navigate safely and autonomously in the campus. We have validated the implemented full autonomy stack on our physical robot. Last but not least, we developed a digital-twin simulation environment for our robot based on the realistic Webots simulator. With the high-fidelity simulation environment, we are able to train learning-based planning algorithms in simulation that could be deployed in the real robot in the future. Specifically, we proposed and implemented a sample-efficient off-policy safe reinforcement learning algorithm CVPO that can learn a safety-aware planner to enable autonomous navigation of different types of agents. We have thoroughly validated the proposed CVPO planner in simulation with simple scenarios, and the results indicate that our method outperforms existing safe reinforcement learning baselines. The corresponding paper will be appeared at the 5th Multidisciplinary Conference on Reinforcement Learning and Decision-Making (RLDM2022). In the future, we will improve our robot platform by incorporating the tracking module, prediction module, and planning module, such that the robot could achieve point to point autonomous navigation tasks around the CMU campus. We will also develop an efficient data collection system to monitor the sidewalks’ status with the autonomy stack of our robot. The objective of task 5 is to develop tools that enable us to analyze the impact of PDD on different communities. During this project, we have gained understanding on how uncertainty in our estimation of various forms of sufficient statistics (e.g., number of users in each neighborhood) may lead to disparate impacts on different communities. The preliminary results in our working paper shed light on how data-driven allocation algorithms should take into account that inherent noise in the underlying data collection process. The objective of task 6 is to develop mechanisms to mitigate PDD’s disparate impact and promote equity. During this project, we have developed a causal inference technique motivated by this research question. In particular, our algorithm leverages the technique of instrumental variable regression, which allows us to estimate the treatment effect of an intervention (e.g., the slowdown of traffic due to the deployment of PDD technologies) on communities of different types, due to their different demographics, levels of density, and infrastructures. This will form the basis of how we design mechanisms to design mechanisms or interventions to mitigate the undesirable effects of PDD on different communities. Impacts We gave a demo to a group of legislators including four senators and explained the techonology to them. We havd submitted a few publications to top venues. Other Website: https://safeai-lab.github.io/MobileRobot.html Outcomes New Partners no Issues no