iiinova — where your digital self keeps evolving

Cathy Wu @cathywu.bsky.social · 27/03/2026

If pure AI or pure search isn't working for you, consider learning-guided optimization! This approach combines the best of both worlds; our most recent work demonstrates this for coordinating agents in warehouse robotics.

140

Cathy Wu @cathywu.bsky.social · 23/02/2026

What's an energy-efficient way to shovel snow? Introducing eco-shoveling*, a two-stage technique to handle snow accumulation, while conserving energy in preparation for more shoveling later... Built to withstand three winter storms.

110

Cathy Wu @cathywu.bsky.social · 19/02/2026

Have you noticed how navigation apps include walking & waiting for public transit, but excludes parking & walking for driving? After being late a few times 😅, we finally did. We got curious: what if these apps account for parking?

67934

Cathy Wu @cathywu.bsky.social · 11/08/2025

This project was 4 years in the making and it's finally out! We found that controlling vehicle speeds to mitigate traffic across a city can cut carbon emissions between 11 and 22 percent. To do this, we used deep reinforcement learning to optimize one million eco-driving scenarios. 🚗🤖🧠

15217

Cathy Wu @cathywu.bsky.social · 02/06/2025

This is the way to cross a freeway. 😍 Breathtaking walkway across SR-520 in Redmond, WA at Microsoft Campus🚶‍♀️🚲🛣️

Beautiful pavilion-topped pedestrian walkway and bike path along 40th street across SR-520 in Redmond, WA

160

Cathy Wu @cathywu.bsky.social · 26/05/2025

📣 I will be giving two talks this week at the TRB DATA conference in Seattle, which explores the intersection of transportation, artificial Intelligence, and data analysis. I love these practitioner + researcher conferences. 🤖🧠👇

Flyer announcing Prof. Cathy Wu from MIT presenting at the TRB Conference on Data and AI for Transportation Advancement. Two sessions are highlighted: one on May 28 about accelerating transportation research transparency practices, and another on May 29 about mitigating metropolitan carbon emissions using dynamic eco-driving.

164

Cathy Wu @cathywu.bsky.social · 25/05/2025

📣 #RERITE 's first conference presentation will take place tomorrow at the Transportation Research Symposium at Rotterdam in The Netherlands! 🚂🛣️ We use Large Language Models (LLMs) to measure the state of data & code availability in transportation research. Join to learn: 👇

Flyer announcing a presentation at the Transportation Research Symposium 2025, scheduled for Monday, May 26, 13:15-15:15 in Rotterdam, The Netherlands. Session topic is 'Traffic control in the era of new technologies,' with the presentation titled 'Towards Accelerating Transportation Research: Measuring the State of Transparency Practices.' Multiple authors from various international institutions are listed, including Ruth Lu, an undergraduate researcher at MIT, whose photo is featured prominently. A QR code and the website rerite.org are displayed in the upper right corner. The flyer background is orange with black and white text.

152

Cathy Wu @cathywu.bsky.social · 24/05/2025

What an utter delight to be taking Sound Transit’s new light rail line!!! (2 Line) 🚉 🌆 #transit

Light rail train platform at Downtown Bellevue towards Downtown Redmond

080

Cathy Wu @cathywu.bsky.social · 16/05/2025

Taizhou's industrial zones www.etaizhou.gov.cn/industrialpa...

Aerial panoramic view of a sprawling industrial park: rows of blue-roofed warehouses and factory buildings stretch toward a wide river, where container cranes line the distant shoreline. Farmland and access roads frame the scene beneath a clear, bright blue sky.

010

Cathy Wu @cathywu.bsky.social · 10/05/2025

5GAA's map of V2X use cases towards cooperative driving Courtesy CCAT www.youtube.com/watch?v=fn_V...

A diagram illustrating the "Evolution of V2X use cases towards cooperative driving," categorizing advancements into three main groups: Automated Driving, Advanced Safety, and Traffic Efficiency/Safety. These groups progressively move through stages labeled Connected Driver Warning, Connected Driver Assistance, and Connected Cooperative Driving. Examples of use cases include Automated Valet Parking, Cooperative Adaptive Cruise Control, Dynamic Intersection Management, Tele-Operated Driving for Automated Vehicles, and Hazard Information Collection and Sharing. Icons depict vehicles interacting via wireless communication, sensors, cameras, and various road safety scenarios.

120

Cathy Wu @cathywu.bsky.social · 05/05/2025

Excited to announce IntersectionZoo, a benchmark that uses a real-world traffic problem to test generalization progress in deep reinforcement learning, particularly multi-agent contextual RL. 🤖🧠 MIT News coverage: news.mit.edu/2025/new-too... Benchmark: intersectionzoo-docs.readthedocs.io

In building IntersectionZoo, three main logical layers are used to abstract functionality as shown in this figure. In the traffic scenario modeling layer, we first build data-driven simulation environments of signalized intersections and then use them to build traffic scenarios at those intersections. Concretely, an intersection is first defined by factors such as lane lengths, lane counts, road grades, turn lane configuration, and speed limit of each approach. Then, vehicle type, age, and fuel type distributions are used with appropriate traffic flow rates and human driven vehicle behaviors to define a realistic traffic flow. Each intersection is then used to define traffic scenarios by further assigning representative atmospheric temperature and humidity values based on the season. Further scenario variations can be achieved by changing the eco-driving adoption level (0%-100%).

Once traffic scenarios are modeled, the CMDP modeling layer is used to contain them as Conextual Markov Decision Process (CMDP) and define the state, action, reward to be used with the reinforcement learning algorithms. CMDP is used to model problem variations. In IntersectionZoo, each city is modeled as a CMDP with each traffic scenario stemming from the city as a problem variation. In a CMDP, each problem variation is an MDP defined by a problem variation context and called a context-MDP. In eco-driving, those context-MDPs are defined by the following state, action and reward functions and formulate as a multi-agent control problem.

In user configuration layer, we provide users the flexibility to configure their experimental setup.

An example of cooperative eco-driving at signalized intersections where CVs are controlled to minimize the total exhaust emissions of the fleet (both CVs and HDVs) while minimizing the impact on the travel time of each vehicle. While CVs are controllable by some defined eco-driving strategy, HDVs are driven by humans and can not be controlled. However, CVs implicitly control HDVs through car-following dynamics and by forming locally cooperative teams for better system control (by controlling the oppotunities of HDVs to overake CVs).

we show the intersection distribution of each city and the distributions of intersection features of each city. The full dataset of 16,334 intersections in compliance with SUMO simulator can be found in the [here](https://drive.google.com/drive/folders/1y3W83MPfnt9mSFGbg8L9TLHTXElXvcHs?usp=sharing).

1183

Cathy Wu @cathywu.bsky.social · 06/01/2025

Find us at the TRB Annual Meeting 2025 this week in DC! PhD student Shreyaa Raghavan and I will share recent developments in reproducibility and new highway traffic research. Hope to see you there! #TRBAM

Flyer containing event details.

Program links:
- https://annualmeeting.mytrb.org/OnlineProgram/Details/22044
- https://annualmeeting.mytrb.org/OnlineProgram/Details/22715
- https://i24motion.org/news/join-us-trb-2025-motion-workshop-and-reception

141

Cathy Wu @cathywu.bsky.social · 05/01/2025

Reproducible research and open science at #TRBAM! Stop by and thank the authors and organizers for contributing their work openly for the community to build upon. Curated by the REproducible Research In Transportation Engineering (RERITE) Working Group Are we missing something? Leave a comment!

Sunday, January 5th
Open-Source Data Repositories and Technologies Workshop
Session: Workshop Location: Convention Center, 150A Time: 9:00 AM - 12:00 PM ET

Monday, January 6th
Understanding Driver–Pedestrian Interactions to Predict Driver Yielding: Naturalistic Open-Source Data Set Collected in Minnesota
Session: Lectern+Poster Location: Convention Center, Salon B Time: 8:00 AM - 9:45 AM ET

Benchmarking the Capabilities of LLMs in Transportation System Engineering
Session: Poster Location: Convention Center, Hall A Time: 1:30 PM - 3:15 PM ET

V2XPnP: A Real-World Sequential Dataset and Fusion Pipeline for Perception and Prediction with Vehicle-to-Everything Communication
Session: Poster Location: Convention Center, Hall A Time: 1:30 PM - 3:15 PM ET

Introducing RERITE: REproducible Research In Transportation Engineering (RERITE) Working Group
Session: ACP50 Committee Meeting Location: Marriot Marquis, M2 Time: 1:30 PM - 5:30 PM ET

Tuesday, January 7th
An Open-Source Framework for Evaluating Cooperative Perception in Urban Areas
Session: Poster Location: Convention Center, Hall A Time: 6:00 PM - 7:30 PM ET

V2X-Real: a Large-scale Dataset for V2X Cooperative Perception
Session: Poster Location: Convention Center, Hall A Time: 6:00 PM - 7:30 PM ET

Imputation and Correction of Traffic Signal Data for Waymo Open Motion Dataset
Session: Poster Location: Convention Center, Hall A Time: 6:00 PM - 7:30 PM ET

Annual Zephyr Reception
Session: Reception Location: Busboys & Poets. Time: 8:00 PM - 10:00 PM ET

Wednesday, January 8th
Building Digital Infrastructure: Scalable and Reproducible Data Integration and Traffic Theory for Congestion Bottleneck Identification and Traffic Management Strategy Evaluation
Session: Poster Location: Convention Center, Hall A Time: 8:00 AM - 9:45 AM ET

I-24 MOTION research workshop
Session: Workshop Location: Open Gov Hub meeting space Time: 2:00 PM - 5:00 PM ET

2101

Cathy Wu @cathywu.bsky.social · 22/12/2024

Yes, ready for some thinking and reading time during the holidays!! Got a list of research questions and books critiquing my field loaded up and ready to go.

4276

Cathy Wu @cathywu.bsky.social · 08/12/2024

12/n This was joint work with my students @jhcho.bsky.social, Vindula, and Sirui. But not only that! This work builds on years of sweat and tears in trying to applying RL, and we hope that this brings some optimism to others as well. n=12

010

Cathy Wu @cathywu.bsky.social · 08/12/2024

9/n We like multi-agent traffic tasks, so we tried those out too. Up to 25x sample efficiency improvements! 🚦🚗 Btw, 25x means training 4 tasks to match the performance of training 100 tasks. I cannot express how excited we are to try this out on more applications.

110

Cathy Wu @cathywu.bsky.social · 08/12/2024

8/n It works surprisingly well! Up to 50x sample efficiency improvements over typical training (brown vs orange, green)! Sequential oracle (pink) chooses the single best training task at each step, given full knowledge. And some heuristic baselines (random, greedy, equidistant).

110

Cathy Wu @cathywu.bsky.social · 08/12/2024

7/n Using this model, we choose training tasks––one at a time––that best improve overall generalization performance across contexts. We leverage ideas of optimism in the face of uncertainty (UCB) and Bayesian optimization to design a suitable acquisition function. Then iterate!

100

Cathy Wu @cathywu.bsky.social · 08/12/2024

6/n The core idea is to explicitly model zero-shot generalization, hence Model-based Transfer Learning (MBTL). We model 1) how well a given RL method would solve a task (Gaussian Process) and 2) the performance loss when applied to another task (linear in context similarity).

100

Cathy Wu @cathywu.bsky.social · 08/12/2024

3/n Surprisingly (?), if we use typical training paradigms, independent training is brittle (orange) and multi-task training is even worse (green). Even on “simple” tasks like Pendulum! I cannot convey how frustrating this is as someone who wants to apply RL to the real-world. 😡

100

Cathy Wu @cathywu.bsky.social · 08/12/2024

2/n OK, let's back up. In deep reinforcement learning (RL), we are almost always interested in solving a range of tasks. As such, we consider contextual RL, where the space of tasks is parameterized.

110

Cathy Wu @cathywu.bsky.social · 08/12/2024

👼 As an applied RL researcher, this is the most optimistic I have been about RL in years. It feels like seeing the light at the end of the tunnel when RL training starts working reliably. Without a ton of compute or tuning. Very excited for what is to come. Here is what we did👇

4304

Cathy Wu @cathywu.bsky.social · 02/12/2024

Got it, that's a great insight. Revised!

030