Attending the Transportation Research Board 2024 Conference
Written by: Nithin Santhanam
Introduction
I recently attended the Transportation Research Board (TRB) Conference from January 7th to January 11th which provided valuable insights into the diverse and dynamic landscape of transportation research. The conference is held in Washington D.C every year and it gathers professionals from every type of sub domain within the transportation field. The aim is to spread awareness, intelligence, and enable collaborative efforts towards common industry goals. The conference primarily consists of presentation sessions, poster sessions, workshops and social events that encourage discussion and the sharing of ideas. It also provides professional networking, job opportunities and exhibition events.
Our Research Interests
I work on the Software-Defined Mobility team at 99P labs at HRI-OH. Our work focuses on understanding mobility challenges and finding innovative approaches and solutions for what we see as critical issues in future mobility. For this reason, it becomes important to understand the transportation domain from policy, commercial, economic, environmental, ethical and many other perspectives. We also want to be able to represent our work in the same terminology and standards that are accepted in the research community.
For those readers who may be unfamiliar, software-defined mobility has a goal of increasing adaptability and efficiency in mobility systems, software-defined mobility refers to exploring software-driven transformations by prioritizing dynamic software control rather than relying solely on traditional hardware configurations. This involves facilitating real-time updates, ensuring connectivity across networks, and providing customized user-focused solutions aimed at revolutionizing mobility.
Our team focuses on a variety of different research areas pertaining to data, data engineering and software development for transportation use cases. My personal area of work relates to using enormous amounts of data for simulations, user behavior understanding and product performance.
My Topics of Interest:
- Traffic Flow Theory
- Network Modeling and Control
- Simulation Methodology and Approaches
- Digital Twin Approaches
- Computer Vision use cases at Scale
- AI-Enabled Simulation Research
- Multi-Modal Data Fusion and Synchronization
- Cooperative Networking and Connected Vehicles
- Edge Computing and Work Offloading
- Large Scale Data Collection Architecture
- Computing Platform requirements for Large Scale Simulation
- Behavioral Science Research Practices in the Transportation Field
- The role of Foundation Models and Large Language Models in the Transportation Field
- Useful tools for Data Science in Big Data
My Conference Experience:
I personally viewed >100 different projects ranging from Poster sessions, lectures, and exhibitions. There was constantly something to do throughout the day. Many presenters are looking for future collaborations, advice, funding or inspiration. Individuals looking for employment are able to attend a big career fair where government groups or other companies host booths and accept resumes for applications to their programs. Since this was my first time attending TRB, I attended the “New Attendees” session where various aspects of the conference offerings are discussed. They help initialize the networking process by introducing you to people interested in a similar domain to yours. They also have many small groups of Q&A sessions for any questions you may have.
TRB is meant to create a lasting community of collaboration in the transportation field and they want to influence the future of all domains of interest. You are able to join committees that discuss cutting edge research, new ideas, critique of previous ideas, etc. You are able to join these committees informally and then be brought on as a member if you are nominated by the committee to be more involved. The impression that I got was that if you have an interest in investing time to assist the committee, they will appreciate the interest and effort and include you in future activities or research planning in the committee. These committees are independent of the annual conference itself so they allow you to immerse yourself into small sub-communities that are interested in the same type of work. This can be a great resource for professionals or students to advance their knowledge, networking or professional research being done. Many collaborations and partnerships are created through these committees and they aim to be as inclusive as possible with rotating board members, public meetings, etc.
I attended the conference alone but it would have helped a lot for multiple team members to attend to see what type of inspiration we all found and then compared. Additionally, in the future I should more carefully plan itinerary so that topics of interest with conflicting time slots aren’t as big of an issue. A lot of related areas of research presented at the same time slots and I was forced to choose one or a few. More team members being present would’ve allowed us to potentially split up when necessary.
In order to attend the conference you need to sign up as a free member on their website and then it will give you the option to attend the Conference for any amount of the days you’d like. The conference is a bit pricey and favors members who are students, young professionals, government employees, or employees of larger companies that have budgets for conferences.
TRB is a massive conference and as a result, most of the surrounding hotels in the area get fully booked quickly so I’d advise to make travel arrangements way ahead of time if possible. Washington D.C was very walking friendly and I was able to find a hotel about 6 blocks away from the convention center.
The food was a little expensive and there weren’t a lot of healthy options from what I saw in the convention center. There were a ton of restaurants open in the surrounding area though that would accommodate anything you would want. The conference developed an app that helps figure out any logistics question that you could have while attending, as well. The app even contains a digital history of almost all the work presented at the Conference in case you missed a topic of interest or wanted to understand any research further. Overall, I thought the conference was extremely well organized and well run.
Traffic Flow Theory, Network Control and Simulation
Our team has attempted data driven and agent-based traffic simulation projects. We work with Carnegie Mellon University on Traffic Simulation and their team has experts in the field helping guide the research. We were relatively new to the area and have been trying to learn as we work on the projects. Our role in the projects are to provide business insights, resources and additional technical expertise. Since the research has a business focused origin, I believe that some of our collaborators discussed the topic in more high-level terms in efforts to ease our understanding of the content. As a result, I was familiar with many terms discussed conceptually but I discovered some of the more official terminology used when discussing traffic flow theory and network control.
Macroscopic Fundamental Diagrams
One noticeable trend that I saw from observing presentations at TRB was that simulation research is often discussed by referring to the Macroscopic Fundamental Diagram (MFD). An MFD is a graph relating an average network flow to the average density and is used to evaluate traffic-control. Using this function, one can describe the traffic of an area at a high level and the function will hold to good accuracy if enough observations are present. We would be able to infer the general flow based on simply knowing the vehicle count or density of vehicles at a given link. This was a very helpful way to view system level traffic flows and it would be helpful to use these types of diagrams as general baselines to understanding system level traffic.
The figure below shows the structure of a typical Deep Learning model that uses spatial characteristics along with the Macroscopic Flow Diagram to yield a speed estimation model describing the traffic flow of a system.
During the conference I found myself drifting towards projects that implemented a theory called Perimeter Control. Perimeter Control attempts to regulate the traffic flows between different regions of a road network by coordinating the signal timings at region boundaries with the aim of improving the overall network performance.
I appreciated the idea of separating complex systems into smaller subsystems and making the process more modular. There were also many different approaches to applying this theory of traffic flow that I would like to continue looking into further.
Model Predictive Control
Model Predictive Control (MPC) was a common methodology used in network control problems. For those new to the framework, imagine you are driving a car on a winding mountain road. Model Predictive Control (MPC) can be compared to a skilled driver who, instead of merely reacting to immediate obstacles or turns, takes into account the entire upcoming road and the characteristics of the car. It is a feedback control approach that makes use of an explicit model of the system to predict its future behavior and determine the optimal control inputs.
- System Understanding (Model): The driver has a detailed map of the road and a clear understanding of how the car responds to different inputs like steering, brakes, and acceleration. This map is analogous to the mathematical model used in MPC.
- Looking Ahead (Prediction Horizon): The driver looks not just at the next turn but plans for a certain distance ahead, anticipating upcoming curves, hills, and straight stretches. In MPC, this corresponds to considering a prediction horizon.
- Optimal Decision Making (Optimization): The driver optimizes their speed and steering at each moment to ensure the best overall progress. Similarly, MPC solves an optimization problem to determine the best control inputs over a defined horizon, considering factors like speed, constraints, and objectives.
- Adjusting in Real-Time (Feedback): As the car moves forward, the driver constantly adjusts the plan based on the evolving road conditions and the car’s actual response. MPC, too, updates its control inputs in real-time based on the system’s feedback and changing conditions.
- Adhering to Rules (Constraints): The driver respects speed limits, road signs, and safety rules. In MPC, constraints are considered to ensure that the calculated control inputs adhere to the limitations of the system.
In essence, MPC is like having a foresighted, adaptive driver who not only navigates the immediate road but strategically plans and adjusts to ensure a smooth and optimal journey over the entire course.
Some of the advantages of MPC:
- Able to fully leverage all the system dynamics
- Requires generic constraints
However, the drawbacks to this strategy are:
- Performs an optimization at a designated rolling time horizon which becomes very computationally expensive
- High algorithm complexity
- Requires all the system dynamics to be known initially, alot of parameter coefficients are required
Many of the projects aimed to simplify MPC or find alternatives that used MPC principles but that can be applied in real-time. For large scale macroscopic simulations, MPC seems like a great choice of methodology to use in order to intelligently and dynamically optimize a network.
Our team has had access to large amounts of telematics data and we’ve had the goal of inputting some of this data into traffic flow simulation models. We have typically used data driven approaches that utilize Origin Destination pairs to calibrate the model. One interesting model that I thought would be promising for our work was the Spatio-Temporal Graph Neural Network. The figure below describes the framework of the STGNN.
Through our research, there was sort of an implicit understanding of the ideas behind what MFD convey but we never explicitly used MFD to discuss our research. A large part of Simulation research is understanding the general behaviors of a system that are dependent on important network factors. Representing complex systems as simpler functions is an extremely powerful tool in the simulation world that enables the approximation of systems in a simpler and less computationally intensive manner.
The figure below shows a project that worked to model traffic flow and it illustrates the inferred power consumption implications on the state electric grid based on forecasted Electric Vehicle adoption. We are very interested in projects like these, and traffic flow is very powerful piece of knowledge that enables other research.
Traffic Simulation Using Reinforcement Learning
My team focuses on the use of large data sources and deriving value for different use cases. As such, we tend to favor data-driven approaches and the corresponding modeling methodologies one would use. Many of the research presentations used Reinforcement Learning (RL) models. There have been successful applications of Deep Reinforcement Learning that have led to positive effects in traffic congestion problems. Some examples of DRL being applied to traffic use cases:
Intuitively, I like the idea of RL models because they work by representing agent behaviors through understandable reward functions. If you can understand the reward-cost relationships from data, transfer learning or domain knowledge, you can represent very complex systems without needing a ton of data. Within transportation use cases, transfer learning has a lot of potential to be applicable to many projects. Many of the core behaviors of mobility can be described similarly and instead of retraining every agent over and over, we can use transfer learning to provide a smarter starting point for training. Since training large models is a cumbersome task for most, transfer learning and the transferability of models has become a larger interest to the Data Science community.
Some of the Different Reinforcement Learning Techniques seen at TRB:
- Q-learning
- Markov Decision Process (MDP)
- Dynamic Programming
- Policy Iteration
- Value iteration
The figure below shows an example of application of a Deep Q-Learning Model used for perimeter control.
Digital Twin, Mapping and Environment Recreation
The exhibition room was an informative experience to see how companies are using transportation inspirations to generate real business solutions and how they choose to market the solutions. One of the most striking things that I noticed from TRB was that a huge percentage of the products and services that I saw were attempting to solve the problem of digital representations of real-life problems. With various methodologies, they all attempted to use sensors such as LIDAR, Laser, Camera, RADAR, etc. to recreate environments as a digital representation that would enable other work to be done.
One product at the exhibition used Photogrammetry, which I hadn’t seen a lot before. Photogrammetry offers an easy to use, relatively low cost and low hardware requirement method to digitally represent 3D environments. However, the accuracy is a lot more variable depending on the operator, the color contrast, the picture quality and often depend on using drones which are difficult to legally operate indoors.
Many products aimed to identify safety hazards such as cracks in roads or railways by using these sensors to recreate the target as a digital twin and then analyze that digital twin through various quality tests.
At 99P Labs we have a digital Twin team that works to model various transportation topics of interest such as battery life, manufacturing robots, traffic flow, energy consumption, etc. We have been working to deploy large scalable digital twins on our data platform. Particularly with 3-Dimensional scene reconstruction, our team has often gone back and forth between what type of sensor suites we believe would be ideal for use cases. This can be due to the sensor cost or effectiveness themselves, the cost of data storage, ease of use, and balancing redundancy vs. reliability. It was interesting to see the variety of different sensor suites used based on the priorities from all the products I saw. Every company is also constantly trying to find the balance between these different factors and there isn’t a clear cut correct answer.
I would like our team to dedicate more time to the field of 3D reconstruction. It seems apparent that there are an enormous amount of products, business models and transportation use cases centered around the methodologies and technologies. Additionally, because of it’s importance, companies are constantly trying to find the most ideal tradeoff optimizations that fit their specific use case and I believe this could be a very exploitable area for our team.
Three-dimensional (3D) reconstruction using sensor data in the transportation industry faces several challenges, primarily due to the dynamic and complex nature of the environments in which these systems operate. Here are some of the key challenges:
Data Fusion and Integration:
- Multimodal Sensor Integration: Transportation systems often use a combination of sensors such as LiDAR, radar, cameras, and GPS. Integrating data from these diverse sources while maintaining accuracy and consistency is a significant challenge.
- Synchronization: Ensuring precise synchronization among different sensors is crucial for accurate 3D reconstruction. Timing misalignments can lead to errors in the reconstructed scene.
Dynamic Environments:
- Moving Objects: Vehicles, pedestrians, and other dynamic elements in the environment can create challenges. Tracking and reconstructing moving objects accurately without introducing errors are complex tasks.
- Scene Changes: Rapid changes in the environment, such as sudden lane shifts, road closures, or construction, require real-time adaptation of the reconstruction system.
Accuracy and Precision:
- Sensor Calibration: Accurate calibration of sensors is essential for precise 3D reconstruction. Calibration errors can result in misalignment and distortions in the reconstructed scene.
- Noise and Distortions: Sensor noise, occlusions, and environmental conditions (e.g., rain, fog) can introduce distortions and reduce the overall accuracy of the reconstructed 3D model.
Computational Complexity:
- Real-time Processing: Transportation systems often require real-time or near-real-time processing for timely decision-making. The computational demands of processing large volumes of sensor data and reconstructing 3D scenes in real-time can be challenging.
Scalability and Robustness:
- Large-Scale Environments: Handling large-scale environments, such as urban areas or highway networks, poses scalability challenges. The system must efficiently process data from extensive regions without compromising performance.
- Robustness to Variability: The system needs to be robust against variations in lighting conditions, weather, and different types of road infrastructure.
Privacy and Security:
- Data Privacy: As sensor data may include information about the surroundings and potentially identifiable objects, ensuring privacy in data collection and processing is a critical concern.
- Security: Protecting sensor data from unauthorized access and manipulation is crucial to maintain the integrity of the reconstructed 3D information.
Regulatory and Standards Compliance:
- Compliance: Meeting regulatory requirements and adhering to industry standards is essential for the deployment of 3D reconstruction systems in transportation. Ensuring interoperability with existing infrastructure is also a consideration.
Maintanence:
- Cost: In order to maintain these systems companies will need methodologies that scale well relative to cost to justify the research or implementation
- Monitoring: Ability to monitor performance in real time and providing a way to supervise some of these processes
- Expertise: Teams with specialties or wide range or domain expertise to enable the effective implementation, maintenance and improvements
Addressing these challenges requires a multidisciplinary approach involving expertise in computer vision, sensor technologies, robotics, signal processing, and real-time systems. Ongoing advancements in these fields contribute to the development of more robust and reliable 3D reconstruction solutions for the transportation industry.
I think we will need to develop a more comprehensive theory behind our digital twin representations and have the choices me make regarding technology and methodology reflect that philosophy so that there is more logical consistency. 2-Dimensional LIDAR is an example of a sensor used in previous projects due to the low cost, general versatility and ease of use, but there are no actual use cases where I think 2D LIDAR would be useful than other sensor choices due to the limited granularity and lack of context. I think RGBD cameras provide a lot of information at a cheap cost and I’d also like to work more with 3-Dimensional LIDAR or laser technology. What will we prioritize and when? Quality of Service? Accuracy? Cost? Lightweight? In the future I’d like our team to create a more consistent angle towards these projects.
Cooperative and Connected Vehicle Modeling
Cooperative and Connected Vehicle multi-modal models were a large theme at this year’s conference. There is a concerted focus on fusion of multiple heterogenous data sources. Many of the projects also focused on cooperative motion planning and cooperative network control to optimize traffic. Cooperative planning and modeling are a very effective way of resolving traffic because you’re able to account for the system equilibrium dynamics as opposed to specific vehicle self-interest. Many papers focused on simulation scenarios where both the system equilibrium and user equilibrium are optimized to meet certain system designs.
The Figure below shows an example of a proposed Human Driven and Connected Vehicle Mixed traffic model:
The Figure below shows an example how one study wanted to use connected vehicle data to update the entire system state to enable collective awareness of every road link’s vehicle flow. It aimed to compute path planning then update the link attributes and then serially update the system level attributes accordingly.
Our team has historically focused on big data platform architectures. Cooperative and Connected Vehicle models typically involve the collection of many different sources of data, several different sub-models being run, a much larger model being run and then an enormous iterative network optimization process running on top of that to reduce the overall system error. Running this type of model on a scalable cloud platform can be very valuable to increase the accuracy and speed of training the model. A common theme amongst most of the discussions was the ability to accommodate the data and computation requirements for very sophisticated models.
I believe with a little more understanding of the core computation principles in creating large network models, our team will be able to leverage our DevOps experience and create a robust environment to train and deploy these models. Additionally, we have experience developing API’s and applications and with a better understanding of our desired use case, I think we can leverage our experience to create a valuable user facing tool that will give access to models such as the type described. We could run “what-if” scenarios that allow a user to calibrate the model to different specifications. We can assess the overall computation requirements and assist Honda in benchmarking various use cases. We can also give a better understanding on the latency requirements and types of communication technology required for effective connected models. Our team has been interested in the field of Smart Cities and using infrastructure to enable more convenient or effective features. Edge computing is another field that is very relevant to the connected technology space and our hope is to provide the environment where we can create simulations to better understand all of these topics.
Human Computer Interaction Research
TRB has many sessions relating to understanding human behavior and creating use cases based on having better understandings of the choices people make. 99P Labs also has a Human Computer Interaction team and it is a goal for us to combine some of our software use cases, tools and practices with our HCI team in an attempt to apply our research to human behavior use cases.
I saw a lot of research pertaining to the creation of business models that enable more effective transportation. I saw presentations on how to conduct Behavioral Studies related to transportation topics. I also saw innovative Behavioral Study Projects that were tested through simulation.
Some Interesting Projects:
Investigating Mode Choice Preferences in a Tradable Mobility Credit Scheme:
Analyzing Travel Behavior and Mode Choice of Underrepresented Youth for Out-of-School-Time Activities Using Discrete Choice Modeling Techniques
CAN-based Naturalistic Driving Data-Collection at Scale
Modeling and Managing Parking and Vehicle Sharing Choices with Autonomous Vehicles
Designing the Publicly-Owned Centralized Platform for Ridehailing Services with Shared Automated Vehicles
Closing Thoughts
I had an incredible time attending this conference and it was very informative for understanding some of the relevant transportation trends and cutting-edge research in fields that my team is interested in. I am hoping to attend this conference again and I would encourage anyone in the field to attend at least once.
I forget sometimes how many sub domains exist inside the enormous umbrella of Transportation Research. As a team member it is easy to get caught up in your specific research area and your specific way of viewing your work. The work presented at this conference was really inspiring for myself as a young professional in the field. It was very clear to see how large of a community works in the same field that I do, and that for many of the complex problems we need to use each other’s knowledge to pursue the most innovative solutions. I think attending TRB changed how I would communicate my work quite a bit. There are commonly understood concepts, terms, methodologies that researchers use in discussion in hopes to continue building on the work and I think that this is incredibly valuable. Many problems a company faces in the transportation domain are not unique to that company. Understanding how others have approached the problem, the advantages, disadvantages, the iterations of work and the most useful findings is critical to continuing to push the envelope. I took away a lot of ideas, inspiration and connections through attending TRB 2024.
This post was meant to be a broad overview of my takeaways from attending TRB and was heavily focused on topics that I found interesting or valuable. I’d encourage anyone to explore the numerous topics that were not mentioned or only briefly touched upon in the post. I know that as a team we are definitely going to dive into several of the topics discussed with a lot more depth and specificity. I look forward to posting about follow up work that we embark on!
If you found any of the topics interesting and wish to stay connected. We invite you to explore more perspectives by subscribing to the 99P Labs’ blog, connecting with us on LinkedIn, or reaching out to collaborate. Through ongoing dialogue, we can drive technology innovation responsibly — aligning incredible capabilities to the experiences that matter most. We see a bright future ahead when technological hype gives way to practical human priorities, and we welcome you to realize it with us.
[Follow Us on Medium]
[Follow Us On LinkedIn]
[Check Out Our Past Engagements]
