Sharing economy in the era of full automation: Evidence from autonomous vehicle on-demand mobility services

When self-driving cars become common, owners can rent them out to ride-hailing apps when they aren't using them. This process, known as AV crowdsourcing, turns private vehicles into a flexible, on-demand fleet of robots. However, coordinating these "robotic assets" is a massive logistical puzzle. You must balance the owner's desire for privacy with the platform's need for reliable service.

A new study from researchers at EPFL explores how to solve this coordination problem. The authors propose a mathematical framework to help mobility platforms decide exactly how much to pay owners and how to dispatch vehicles. This helps minimize costs while keeping passengers moving. The core insight is that service quality depends on how well the timing of owner availability matches passenger demand.

The friction of mismatched rhythms

Current models for on-demand mobility (MoD) services, like Uber or Lyft, typically assume the platform owns or controls the entire fleet. This simplifies things. If a driver is needed in a certain area, the platform sends them there. But in a world of autonomous vehicles (AVs), a significant portion of the supply might be privately owned.

The authors point out that existing research often uses "aggregate models." These are simplified math frameworks that look at total numbers rather than specific times and places. This fails to account for the messy reality of human life. Unlike a company-owned robot, a private AV owner has a schedule. They might need their car for a commute at 8:00 AM or a grocery run at 5:00 PM.

If a platform tries to rent a car right when the owner needs it, the system breaks. Previous studies have struggled to capture this "spatiotemporal heterogeneity" (the way supply and demand shift across both space and time simultaneously). Without accounting for these shifts, a platform might overpromise service in areas where owners are actually busy using their own cars.

Coordinating the robotic fleet

To bridge this gap, the authors developed a time-expanded network flow model. Think of this as a high-resolution digital map of a city. It doesn't just show roads, but also tracks the movement of every vehicle through every minute of the day.

The authors reformulate this complex problem into a convex quadratic program (a type of mathematical optimization that ensures a single best solution can be found efficiently). This allows the platform to jointly optimize two main levers:

1. Dynamic Pricing: The platform sets region-specific rental prices. This isn't just about making money. It is a signaling mechanism. By raising prices in certain areas or times, the platform incentivizes owners whose "opportunity cost" (the minimum compensation they require to part with their car) is met.
2. Centralized Dispatching: The platform manages the movement of both its own fleet and the crowdsourced private AVs. This includes managing "empty vehicle flow"—the movement of cars to where they will be needed next.

As shown in, the model treats the private AV as a resource that switches between "owner usage" and "MoD service." The platform acts as the conductor for both.

Lessons from the streets of Chicago

The researchers validated their model using a case study built on real-world data from Chicago. This included ride-hailing trips and household travel surveys. They categorized owners into eight distinct "clusters" based on their habits. These ranged from "home-oriented" users to "intra-city commuters" [Table 2].

The paper reports several key findings regarding service performance:

• Low Participation, High Impact: Across most owner clusters, the daily average rental participation rate was relatively low. It ranged from roughly 1% to 2% . Despite this, these vehicles are critical for filling gaps in service.
• The Importance of Diversity: The study finds that the most efficient service comes from a diverse mix of owners. Specifically, a combination of roughly 50% commuters and 50% non-commuters yielded the lowest platform costs .
• Spatial Concentration: Crowdsourced AVs tend to concentrate in central commercial areas. Meanwhile, the platform's own fleet acts as a "safety net." It spreads out to cover residential and peripheral regions .

One of the most striking results concerns the "reliability buffer" (the time an owner reserves to ensure they have their car when they need it). The authors find that a larger buffer significantly degrades service quality . If owners are highly risk-averse, the platform loses its most valuable flexible supply.

Limits of the mathematical lens

While the model is robust, the authors acknowledge several simplifying assumptions. First, the model assumes that the opportunity costs of owners follow a uniform distribution. In reality, owner preferences might be much more skewed or unpredictable.

Second, the study assumes that the "activity patterns" of owners are fixed. It does not account for the possibility that the crowdsourcing service itself might change human behavior. For example, if an owner realizes they can make significant revenue renting their car, they might change how they travel.

Finally, the model assumes all private AVs are returned to the same region they started in. In a real city, an owner might drive their car to a different neighborhood for an errand. This would create a "drift" in supply that the current model does not fully explore.

The verdict: A blueprint for scaling

Is AV crowdsourcing ready for prime time? Based on this evidence, the answer depends on the diversity of your supply.

The paper demonstrates that AV crowdsourcing is not a magic bullet for all mobility needs. However, it is a powerful tool for managing peak demand if managed correctly. The research suggests that platforms should not just look for more cars. Instead, they should look for different cars. A fleet composed only of commuters will leave massive gaps in service during the midday and late-night hours.

For engineers building the next generation of mobility platforms, the takeaway is clear. Successful scaling will require moving beyond simple aggregate supply models. You must build systems capable of navigating the complex, shifting rhythms of human life to unlock the potential of private robotic assets.

Figures from the paper