Around 30% of a city’s carbon emissions come from transport, with urban deliveries being a major contributor. The rapid growth of e-commerce has led to a surge in light goods vehicles (LGVs) in cities, bringing challenges of increased pollution, congestion, and parking difficulties. By 2030, urban last-mile deliveries are projected to increase both CO₂ emissions and traffic congestion by 30% in the top 100 cities globally.
While cargo bikes have emerged as a promising zero-emission alternative, with studies suggesting they could replace over half of urban van deliveries, logistics operators lack the data and tools to accurately assess their performance in different urban contexts. This creates a significant barrier to their widespread adoption.
Our research project was conducted in collaboration with AI researchers from IIT (Kharagpur 🇮🇳), ITU (Copenhagen 🇩🇰), MIT-IBM (Cambridge, MA, 🇺🇸), in partnership with two expert cargo-bike logistics operators, Pedal Me in London 🇬🇧 and Urbike in Brussels 🇧🇪. The project was supported by a CCAI innovation grant.
We introduce a novel framework for measuring and modelling the performance of delivery vehicles across diverse urban environments. Our approach combines real-world delivery data with advanced machine learning techniques to predict how different vehicles perform in various urban contexts, enabling data-driven decisions about fleet composition and routing.
The efficiency of urban deliveries depends heavily on the local environment – from building density and road networks to the availability of parking. Yet most delivery optimisation models treat cities as uniform spaces, failing to account for how different urban contexts affect vehicle performance.
Our research addresses this gap through three key innovations:
Our research combines three unique datasets:
This combination of datasets allowed us to study patterns of how different vehicles perform across urban contexts. We found that van service times are significantly impacted by urban density, with downtown areas showing service times over two minutes longer than outer urban areas. With delivery drivers making an average of 140 stops per day, this seemingly small difference adds up to over 4.5 hours of additional time per driver per day in dense urban areas. In contrast, cargo bikes demonstrate much less sensitivity to urban context, likely due to their ability to park closer to delivery points.
One of our most significant findings is that cargo bikes maintain consistent service times across different urban environments. The median parking distance for cargo bikes is just 20 meters from the delivery point, compared to van drivers in Seattle spending an average of 138 seconds searching for parking. This suggests cargo bikes could provide more reliable service in dense urban areas where vans struggle with parking and congestion.
Our framework enables several practical applications:
The models we’ve developed can be applied to new urban areas, helping operators evaluate the potential of cargo bikes without costly pilot programs, as well as accelerate the transition to more sustainable urban logistics systems.
While our research demonstrates the potential of machine learning to improve urban logistics, it also highlights areas for future work. We’re particularly interested in expanding our models to account for the role of micro-hubs in enabling cargo bike operations and the impact of new infrastructure on delivery performance.
By continuing to refine these models and gather more data, we can help create more efficient, sustainable, and liveable cities. The code and datasets from this research will be made available to the community to encourage further innovation in this crucial field.
Our work shows that by understanding and predicting how different vehicles perform across urban contexts, we can make significant strides toward more sustainable urban logistics. As cities grapple with the challenges of growing delivery demand and environmental concerns, data-driven approaches like ours will be essential for building more efficient and sustainable delivery systems.
You can read our blog about the project on the Climate Change AI blog.