Rafal Kucharski Lab

We study how people travel. We want to understand their travel demands and behaviour better. We want to see how they use novel mobility services (like ride-pooling or mobility platforms) and how we can design them to better fit the needs of: users, suppliers and sustainable cities. We want to forecast the future of urban mobility with connected autonomous vehicles. We use big empirical datasets, real-world networks, old and new algorithms and models. We simulate, model, analyze, solve problems, propose algorithms to better understand how cities of the future will work. See our works here. The group is led by Rafał Kucharski, associate professor at Group of Machine Learning Research, Faculty of Mathematics and Computer Science at Jagiellonian University in Kraków, Poland. Former PostDoc at TU Delft (prof. Oded Cats), PhD of prof. Guido Gentile (La Sapienza) and alumni of prof. Andrzej Szarata (Cracow University of Technology). In our research we do stuff which can be classified as:

model estimation, optimization, system control, network design;
agent-based simulation, game-theory, network science, stochastic simulation, epidemic modelling;
machine learning, spatial analysis, big data analysis, pattern recognition, unsupervised learning;
behavioural modelling, economic discrete choice models, policy, sustainability.

Currently there is ten of us, working in two different projects, at the modern campus of one of the oldest universities in Europe (est. 1364), Jagiellonian University in Kraków, Poland. We have two major projects:

ERC Starting Grant COeXISTENCE, where we simulate future of cities shared by humans and autonomous vehicles. We use reinforcement learning to optimize joint actions of collaborative machines (cars) and see how it affects the well studied complex social system of urban traffic - will it remain in the Nash Equilibrium? We do not think so, but that’s what we want to demonstrate - stay tuned,
Horizon Europe SUM project - where we apply our in-house ride-pooling algorithms to see the potential of on-demand transit in urban areas of Jerusalem and Kraków
and one already finished: NCN Opus - where we look at the future of ride-pooling and platform services in post-pandemic world.

Feel free to reach to us for a joint seminar, collaboration or vacancies.

drawing drawing drawing drawing drawing

Tweets by Rafal_Lab

selected publications

EJOR

Balancing profit and traveller acceptance in ride-pooling personalised fares

Bujak, Michał, and Kucharski, Rafał

European Journal of Operational Research 2025

Abs HTML

In a ride-pooling system, travellers experience discomfort associated with a detour and a longer travel time, which is compensated with a sharing discount. Most studies assume homogeneous travellers that receive either a flat discount or, in rare cases, a proportional to the inconvenience. This simplified approach offers inaccurate results and leads to an underperforming service when tested against diverse and natural human behaviour. We improve the standard approach on two bases. First, we propose a stochastic setting, where we leverage the population distribution of behavioural traits to determine the acceptance probability. Second, we personalise fares. Each traveller receives a sharing discount based on their contribution to the system such that the operator maximises his expected profitability. In the study, we rigorously prove that the discount optimisation problem can be decomposed. We optimise discounts at a ride level to claim the system optimum. An operator, when proposing fares, encounters two counteracting effects. Low fares increase realisation probability while high fares improve profit from a realised ride. In the personalised discount optimisation, we seek the golden mean. Travellers, who are well-aligned and experience minimal discomfort of sharing, are offered higher fares than those who require more incentive to join the service. Unlike in previous methods, our approach naturally balances the travellers satisfaction and the profit maximisation. With an experiment set in NYC, we show that this leads to significant improvements over the flat discount baseline: the mileage is reduced by 4.5% and the operator generates more profit per mile (over 20% improvement)
TR:C

MoMaS: Two-sided Mobility Market Simulation Framework for Modeling Platform Growth Trajectories

Ghasemi, Farnoud, and Kucharski, Rafal

Transportation Research Part C: Emerging Technologies 2025

Abs HTML

Mobility platforms such as Uber and DiDi have been introduced in cities worldwide, each demonstrating varying degrees of success, employing diverse strategies, and exerting distinct impacts on urban mobility. We have observed various growth trajectories in two-sided mobility markets and understood the underlying mechanisms. However, to date, a realistic microscopic model of these markets including phenomena such as network effects has been missing. State-of-the-art methods well estimate the macroscopic equilibrium conditions in the market, but struggle to reproduce the individual human behavior behind and complex growth patterns sensitive to platform strategy and policies. To bridge this gap, we introduce the MoMaS (two-sided Mobility Market Simulation) framework to represent growth mechanism in two-sided mobility markets based on the realistic behavior adjustment of drivers and travelers reactive to platform strategy. In the proposed framework, traveler and driver agents learn the platform utility from multiple channels: their own experience, peers’ word-of-mouth, and the platform’s marketing, all-together constituting the agent’s perceived utility of the platform. Each of these channels is modeled and updated by our S-shaped learning model day-to-day which stabilizes, and at the same time, remains sensitive to the system changes. The platform can simulate any strategy on five levers: trip fare, commission rate, discount rate, incentive rate, and marketing. While detailed empirical data and actual strategies for platform growth remain largely unknown, MoMaS allows to reproduce series of plausible growth trajectories that were previously unattainable. The framework facilitates the modeling of individual-level behaviors such as reluctance, neutrality, and loyalty, alongside aggregate-level dynamics like critical mass, bandwagon effects, and both positive and negative cross-side network effects. We illustrate the capabilities of MoMaS through an extensive set of real-world experiments. Our results demonstrate that once the platform acquires critical mass, it triggers a significant positive cross-side network effect, accelerating growth. However, this can be reversed if a negative cross-side network effect is triggered, leading to the collapse of the platform. MoMaS is applicable for real-sized problems and available on public repository along with reproducible experiments.
arXiv

URB - Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles

Akman, Ahmet Onur, Psarou, Anastasia, Hoffmann, Michał, Gorczyca, Łukasz, Kowalski, Łukasz, Gora, Paweł, Jamróz, Grzegorz, and Kucharski, Rafał

arXiv preprint arXiv:2505.17734 2025

Abs HTML

Connected Autonomous Vehicles (CAVs) promise to reduce congestion in future urban networks, potentially by optimizing their routing decisions. Unlike for human drivers, these decisions can be made with collective, data-driven policies, developed by machine learning algorithms. Reinforcement learning (RL) can facilitate the development of such collective routing strategies, yet standardized and realistic benchmarks are missing. To that end, we present \our: Urban Routing Benchmark for RL-equipped Connected Autonomous Vehicles. \our is a comprehensive benchmarking environment that unifies evaluation across 29 real-world traffic networks paired with realistic demand patterns. \our comes with a catalog of predefined tasks, four state-of-the-art multi-agent RL (MARL) algorithm implementations, three baseline methods, domain-specific performance metrics, and a modular configuration scheme. Our results suggest that, despite the lengthy and costly training, state-of-the-art MARL algorithms rarely outperformed humans. Experimental results reported in this paper initiate the first leaderboard for MARL in large-scale urban routing optimization and reveal that current approaches struggle to scale, emphasizing the urgent need for advancements in this domain.
SoftwareX

RouteRL: Multi-agent reinforcement learning framework for urban route choice with autonomous vehicles

Akman, Ahmet Onur, Psarou, Anastasia, Gorczyca, Łukasz, Varga, Zoltán György, Jamróz, Grzegorz, and Kucharski, Rafał

SoftwareX 2025

Abs HTML

RouteRL is a novel framework that integrates multi-agent reinforcement learning (MARL) with a microscopic traffic simulation for the development of efficient collective route choice strategies for autonomous vehicles (AVs). The proposed framework models the daily urban route choices of driver agents of two types: human drivers, emulated using behavioral route choice models, and AVs, modeled as MARL agents optimizing their policies for a predefined objective. RouteRL aims to advance research in MARL, transport modeling, and human–AI interaction for transportation applications. This study presents a technical report on RouteRL, outlines its potential research contributions, and showcases its impact via illustrative examples. Initial findings show that human travel times could increase following the introduction of AVs in urban environments, highlighting the importance of future studies to support efficient urban mobility.
Sci. Rep.

Social implications of coexistence of CAVs and human drivers in the context of route choice

Jamróz, Grzegorz, Akman, Ahmet Onur, Psarou, Anastasia, Varga, Zoltán György, and Kucharski, Rafał

Scientific Reports 2025

Abs HTML

Suppose in a stable urban traffic system populated only by human driven vehicles (HDVs), a given proportion (e.g. 10 %) is replaced by a fleet of Connected and Autonomous Vehicles (CAVs), which share information and pursue a collective goal. Suppose these vehicles are centrally coordinated and differ from HDVs only by their collective capacities allowing them to make more efficient routing decisions before the travel on a given day begins. Suppose there is a choice between two routes and every day each driver makes a decision which route to take. Human drivers maximize their utility. CAVs might optimize different goals, such as the total travel time of the fleet. We show that in this plausible futuristic setting, the strategy CAVs are allowed to adopt may result in human drivers either benefitting or being systematically disadvantaged and urban networks becoming more or less optimal. Consequently, some regulatory measures might become indispensable.