Abstract
We aim to investigate how the sidewalk infrastructure affects Person Delivery Devices (PDDs) and residents. To this end, we plan to collect data of the sidewalk ecosystem in Pittsburgh using a robot and establish a multi-model simulator for PDDs at the city scale. We will analyze PDD’s disparate impact on different communities in simulation and explore pricing mechanisms to mitigate the disparate impact and promote equity. The project is of great interest to both the government and industry.
Description
Motivation and Research Goal
Last month, Pennsylvania legalized autonomous delivery robots aka Personal Delivery Devices (PDD) by classifying them as pedestrians. In this new law (Senate Bill 1199 [1]), PDD is defined as a device manufactured for transporting cargo and goods in a pedestrian area, trafficway, or on the shoulder of a highway or roadway, which should limit speeds to 12 miles per hour in a pedestrian area, or 25 miles per hour on the berm of the roadway with weight less than 550 pounds. Besides Pennsylvania, ten other states including Arizona, Florida, Idaho, Ohio, Tennessee, Texas, Utah, Virginia, Washington, and Wisconsin, have legalized PDDs, which has greatly promoted the deployment of PDDs during the pandemic to transport food and medicines.
Like any transportation device, PDDs rely on corresponding infrastructure facilities to operate properly. Currently, most PPDs run on sidewalks. “Understanding how the deployment of PDDs will affect sidewalk infrastructure and natural human users of the public sidewalk, as well as PDDs general impact on different communities is of great interest and importance to the City of Pittsburgh Department of Mobility and Infrastructure” (see the letter of support). Specifically, the width, height, and flatness of sidewalks greatly affect the speed, decisions, and path-planning of PDDs. City managers also want to know when the sidewalks are mostly used by pedestrians to avoid congestion. In addition, obstacles like garbage bags or falling twigs may force PDDs to detour to the main roads which could confuse cars or cyclists. These factors will not only cause safety and cost concerns for companies when deploying PDDs but also may block residents with poor facilities from timely and economically acquiring food and medicine via PDDs.
Despite the quick arrival of PDDs and the importance of sidewalks, city managers have gotten very limited tools to quantitatively understand how the infrastructure affects PDDs and how PDDs affect sidewalk users and local residents. Recently, New York City has launched an online app to measure the width of its sidewalks [2]. However, knowing the width is far from enough to fully understand the whole sidewalk ecosystem. Further study is urgently needed.
In this research, we aim to develop a comprehensive and efficient way to collect, process, and analyze how the sidewalks may affect the PDDs and how PDDs affect other sidewalk users and local residents. Specifically, we want to answer the following questions: 1) how to efficiently collect data of sidewalks?; 2) how to process, analyze, and visualize the data to understand how it may affect PDDs and vice versa?; 3) how to analyze the impact of PDDs for communities with different infrastructure conditions and sidewalk users? These questions have resonated with the policy analysts of the city of Pittsburgh who showed great interest in participating in this study and Starship, which is the main deployer of PDDs at Pittsburgh.
Proposed Methods
In this research, we plan to use the city of Pittsburgh, PA as a living laboratory and use a PDD developed at CMU to collect data and analyze the ecosystem of PDDs on sidewalks. We. In our prior project supported by Manufacturing Future Initiatives, we have developed two logistic PDDs which have been tested at Mill 19 as indoor autonomous delivery robots. We will upgrade one of them with additional sensors and communication tools so that it could be deployed outdoors.
We will then develop data processing, storage, and analysis tools. The confidence to complete this procedure in time comes from our previous research experience on self-driving cars for scenario recognition with end-to-end Bayesian nonparametric learning. We will develop hierarchical machine learning algorithms to statistically model the variance of sidewalks at different locations and time periods. Our previous work [3] shows the Dirichlet Processes-Hidden Markov Model-Gaussian Processes (DP-HMM-GP) model could be a powerful tool for this task. The DP-HMM-GP method will adaptively adjust the complexity of the statistical sidewalk model and automatically segment and cluster the spatial-temporal raw data. Due to the limited budget, we can only collect data on a few typical sidewalks. We will then use transfer learning methods to generate data in simulation for areas with less collected data. Finally, we will integrate and visualize the information we collected on the traffic-net.org website [4], which we built in a previous project supported by Traffic21.
After data collection and modeling of the sidewalk across the city, we will analyze how the PDD deployments may affect different communities in the city. In particular, we aim to provide economic models to analyze the potential disparate impact of PDD on different sub-populations. For example, if a neighborhood has the lower-quality infrastructure, the corresponding PDD service may incur a higher cost, and so the company may choose not to provide any delivery services in such a neighborhood. Furthermore, the delivery company may choose to deploy PDDs during the hours where some neighborhood has the most pedestrians on the sidewalk. To provide a comprehensive analysis, we will consider a wide range of factors, including but not limited to sidewalk quality, traffic flow during different times of the day, and the availability of curb space for parking. Given our disparate impact analysis, we will also study what type of mechanisms transportation agencies can put in place to promote equity. As an example, we can consider lowering the price of running PDD on certain sidewalks in order to incentivize PDD deployment in underserved neighborhoods. We will draw tools from our prior work [5-8] on pricing mechanisms that balance the objectives of economic efficiency as well as equality of services across different sub-populations. We will also leverage the data from Mobility Data Specification (MDS) [9] to design a refined pricing mechanism that can charge based on the time, duration, and location of the PDD usage.
Previous Experience
The research team is uniquely positioned to accomplish the research aims outlined above due to previous work applying data collection, modeling and simulation tools, artificial intelligence, data mining, operations research, equity analysis, and sensing to improve transportation. Ding Zhao is an expert in autonomous robots and AI safety, and his research studies on how to model the digital twin of transportation and assessment of the safety of autonomous vehicles. Steven Wu is an expert in machine learning, and his research studies how to make machine learning better aligned with societal values, especially privacy and fairness.
Reference:
[1] Regular Session 2019-2020, Senate Bill 1199
https://www.legis.state.pa.us/cfdocs/billinfo/billinfo.cfm?sind=0&syear=2019&body=S&type=B&bn=1199
[2] Sidewalk Widths NYC, https://www.sidewalkwidths.nyc/#13/40.714/-74.005
[3] C. Zhang, J. Zhu, W. Wang and D. Zhao, "A General Framework of Learning Multi-Vehicle Interaction Patterns from Video," 2019 IEEE Intelligent Transportation Systems Conference (ITSC), Auckland, New Zealand, 2019, pp. 4323-4328, doi: 10.1109/ITSC.2019.8917212.
[4] J. Zhu, W. Wang and D. Zhao, "Attempt to Unify Heterogeneous Driving Databases using Traffic Primitives," 2018 21st International Conference on Intelligent Transportation Systems (ITSC), Maui, HI, 2018, pp. 2052-2057, doi: 10.1109/ITSC.2018.8569940.
[5] A. Roth, A. Slivkins, J. Ullman, Z. S. Wu “Multidimensional Dynamic Pricing for Welfare Maximization” ACM Transactions on Economics and Computation, 2020
[6] A. Roth, A. Slivkins, Z. S. Wu “Watch and learn: Optimizing from revealed preferences feedback.” Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing, STOC 2016.
[7] R. Rogers, A. Roth, J. Ullman, Z. S. Wu“Inducing approximately optimal flow using truthful mediators” Proceedings of the sixteenth ACM conference on Economics and computation, 2015
[8] S. Kannan, M. Kearns, J. Morgenstern, M. Pai, A. Roth, R. Vohra, Z. S. Wu “Fairness incentives for myopic agents.” Proceedings of the 2017 ACM Conference on Economics and Computation
[9] Mobility Data Specification.
https://github.com/openmobilityfoundation/mobility-data-specification
Timeline
In order to achieve the goal, the research can be divided into the following tasks:
Task 1: Upgrade the robots and scan the sidewalks
7/1/2021-12/31/2021
Task 2: Develop methods to analyze the sidewalk data and identify typical scenarios
8/30/2022-11/31/2022
Task 3: Generate simulated data for areas with less data
10/1/2022-2/31/2022
Task 4: Analyze the operation influence for different scenarios
11/1/2021-2/30/2022
Task 5: Analyze PDD’s disparate impact on different communities
8/1/2021-4/30/2022
Task 6: Explore pricing mechanisms to mitigate PDD’s disparate impact and promote equity
1/1/2022-6/30/2022
Strategic Description / RD&T
Deployment Plan
Task 1: Upgrade the robots and scan the sidewalks
In this task, our team will develop a mobile robot for sidewalks data collection. By leveraging our team’s experience in building autonomous vehicles, which was supported by Traffic21, we will upgrade our existing four-wheel mobile robot platform for outdoor usage with the latest sensing systems, including stereo cameras, multi-beam LiDARs, high-accuracy GPS, and cellular communication. We will also develop the whole data collection pipeline based on the sensors and the onboard computer, and design a real-time human-machine interaction interface and data visualization system, so that we can facilitate the data collection process and instantly monitor abnormal situations.
With the sidewalks data collection mobile platforms, we will digitalize the physical real-world sidewalks information based on our team’s expertise in machine learning and robot perception. Specifically, we will perform semantic 3D point cloud simultaneous localization and mapping (SLAM) to extract both geometric and semantic representation from the LiDAR sensor, including the 3D map and objects semantic meaning information along the sidewalk. We will explore efficient semantic representations of the collected data by using deep learning perception models, such as multi-object detection and tracking, so that the collected data can be better used for our downstream tasks.
Task 2: Develop methods to analyze the sidewalk data and identify typical scenarios
We will then quantify the navigation difficulty from learned scenarios and patterns as automation complexity. Specifically, we will link local transportation features and mobile robot functionalities using mobile robot behavior as the intermediate node. To this end, this project will develop an automatic way to build a scenario library from rich naturalistic data. We will implement an HDP-DP-HMM algorithm to analyze the data. Furthermore, we will study the barycenter of the statistical model to unveil the typical and abnormal scenarios. We will manually check the extracted typical scenarios to evaluate the quality of data processing.
Task 3: Generate simulated data for areas with less data
While real data provide valuable information about the sidewalk ecosystem, it is impractical to collect data at a city scale in this project. We will leverage on our recent research on transfer learning to generate artificial data for areas that we do not collect enough data. It is crucial that the simulated data can correctly represent the real world. Leveraging the distribution alignment methods, we can obtain an optimal measure that describes the difference between different traffic patterns. Thus we can regularize the simulation to be close to the real world.
Task 4: Analyze the operation influence for PDDs in different scenarios
In this task, we will perform a systematic analysis based on the digital twin system developed by Task 2 and Task 3. We aim to build a statistical model to describe the condition of sidewalks in different areas of the city and predict their potential impact on the operations of PDDs. Given the desired metrics, such as PDDs operation time, we will use optimization tools to obtain the optimal PDDs operation schedule and path according to different scenarios.
Task 5: Analyze PDD’s disparate impact on different communities
In this task, we will leverage the data we collect using our robots to analyze how the PDD deployments may affect different neighborhoods and the relevant communities. In order to examine different factors about the sidewalks across the city, our first step is to develop an economic model on the deployment cost that captures a wide range of factors, including sidewalk quality, traffic flow during different times of the day, and the availability of curb space for parking. Secondly, we will study how the heterogeneous deployment costs across different communities influence the delivery company’s deployment policy. In particular, we will analyze if the delivery company will choose to deploy PDDs in certain neighborhoods and during what time does PDD deployment take place in each neighborhood.
Task 6: Pricing mechanism design to mitigate PDD’s disparate impact and promote equity
Given our disparate impact analysis, we will also study what type of mechanisms government transportation agencies can put in place to mitigate such disparate impact and promote equity. We will mainly focus on the design of pricing mechanisms that can influence the deployment of PDDs across the city. We will develop refined pricing strategies that can leverage the information one could collect via Mobility Data Specification (MDS) that includes deployment time, duration, and location of PDDs. In particular, the mechanism can lower the price of running PDD in underserved neighborhoods in order to incentivize PDD deployment. Moreover, the pricing strategy can influence PDD deployment such that PDDs run during periods with less pedestrian traffic. As a concrete goal, we will extend our prior work on congestion pricing [7] that optimizes economic efficiency to develop pricing strategies that are fairness-aware. In particular, our prior work on mechanisms that incentivize fair decision making [8] provides a promising set of tools. Lastly, we will also leverage techniques from our prior work [5,6] to develop dynamic pricing strategies that adaptively set the prices based on observed traffic flow in different neighborhoods.
Expected Outcomes/Impacts
1. Novel tools and approaches to comprehensively collect, process, analyze, and visualize how sidewalks affect the operation of PDDs
2. City level analysis in simulation on how the sidewalks and natural human activities will affect the PDDs and vice versa.
3. Provide a rigorous framework for examining the fairness and equity issues due to the use of PDDs
4. Provide guidance on pricing mechanisms on PDD usage based on MDS data
Expected Outputs
TRID
Individuals Involved
| Email |
Name |
Affiliation |
Role |
Position |
| bmdayak@exchange.cs.cmu.edu |
Dayak, Bernadette |
Carnegie Mellon University |
Other |
Other |
| nancyi@andrew.cmu.edu |
Igoe, Nancy |
Carnegie Mellon University |
Other |
Other |
| zstevenwu@cmu.edu |
Wu, Steven |
Carnegie Mellon Institute for Software Research |
Co-PI |
Faculty - Untenured, Tenure Track |
| dingzhao@cmu.edu |
Zhao, Ding |
Carnegie Mellon School of Engineering |
PI |
Faculty - Untenured, Tenure Track |
Budget
Amount of UTC Funds Awarded
$99997.00
Total Project Budget (from all funding sources)
$199997.00
Documents
| Type |
Name |
Uploaded |
| Data Management Plan |
UTC_Sidewalk_Data_Management_Plan.docx |
Dec. 17, 2020, 8:14 p.m. |
| Presentation |
Towards_a_Smart_Safe_and_Sustainable_Sidewalk_A_Quantitative_Analysis_on_How_Sidewalk_Infrastructure_Affect_Personal_Delivery_Devices.pptx |
March 18, 2021, 2:03 p.m. |
| Publication |
Stateful Strategic Regression |
Oct. 1, 2021, 4:34 a.m. |
| Progress Report |
365_Progress_Report_2021-09-30 |
Oct. 1, 2021, 4:34 a.m. |
| Progress Report |
365_Progress_Report_2022-03-30 |
March 30, 2022, 7:23 p.m. |
| Final Report |
365_-_Final_report.pdf |
Aug. 3, 2022, 4:52 a.m. |
Match Sources
No match sources!
Partners
| Name |
Type |
| City of Pittsburgh |
Deployment Partner Deployment Partner |
| Starship |
Deployment Partner Deployment Partner |
