Navigation that factors in parking: an MIT model shows how to reduce delays, congestion, and emissions
Drivers rely on navigation apps every day to estimate how long it will take to reach their destination. But in practice it often happens that the “estimated time of arrival” turns out to be unrealistic because the last few minutes (or tens of minutes) are spent circling in search of a parking space. Precisely that “hidden” part of the trip — driving to parking, searching for a space, and walking to the destination — most of today’s routing systems do not include in their calculation, even though in city centers it is decisive for the true overall arrival.
Why the problem is bigger than driver frustration
Underestimating the time needed to park is not just a matter of nerves and missed meetings. When many drivers simultaneously circle around looking for a free spot, local congestion grows, fuel consumption and emissions increase, and the streets around popular zones suffer the most — business districts, cultural venues, and major sports or entertainment events. This also creates a broader effect: if an app systematically “prettifies” the duration of a car trip, users may wrongly conclude that driving is faster than public transport, a bicycle, or combined options. In reality, especially at peak times, the difference can reverse as soon as realistic parking and walking time are included.
In the traffic literature this phenomenon is described as “cruising for parking” — circling in search of parking. A review of earlier empirical studies suggests that searching for a parking space can take several to a dozen minutes, and the share of vehicles cruising varies depending on the city, time of day, and level of congestion. In the summary of Donald Shoup’s classic review, a range from 8 to 74 percent of traffic is cited that, under certain conditions, can be attributable to cruising in search of parking, illustrating how large that “invisible” traffic can be when parking supply does not keep up with demand.
The MIT approach: it doesn’t guide the driver to an address, but to the “best” parking
A team of researchers from the Massachusetts Institute of Technology (MIT) proposed a model that puts the probability of successful parking at the center of navigation. Instead of always sending the user to the closest parking next to the destination, the system looks at a set of public parking facilities in the vicinity and estimates the total expected “door-to-door” time: driving from the origin to the selected parking, the likely time to find a space, and walking from the parking to the goal. The key idea is that “closest” is not necessarily “fastest” when demand is high, because a parking facility right next to the destination may have a very low chance of having a free space at that moment.
In the paper “Probability-Aware Parking Selection”, whose preprint was published on arXiv on 2 January 2026, authors Cameron Hickert, Sirui Li, Zhengbing He, and Cathy Wu formulate the problem by introducing a probabilistic layer — the probability that the user will find a space at a particular parking facility — alongside distance and travel time. The model then seeks a strategy that minimizes the expected arrival time, taking into account failure scenarios as well: what if the driver arrives at the “ideal” parking facility, but the spaces are full? Instead of improvising on the spot, the algorithm computes in advance what the next best options are, how far they are, and what the risk is that they too will be full.
Dynamic programming and “traffic as a multi-player game”
The technical backbone of the approach is a framework based on dynamic programming, which “works backward” from favorable outcomes and calculates optimal decisions under uncertainty. What makes the model closer to a real city is the fact that it does not assume the user is the only one looking for parking. In reality, many drivers act simultaneously, and their decisions change the probability of success for everyone else.
The example is simple: another driver can arrive a second earlier and take the last free space, or can first try one parking facility and, if unsuccessful, “spill over” to another — perhaps precisely the one the system recommended to the user. Exactly such spillover effects, which in practice are common in cities with a series of smaller garages and parking lots at short distances, can change the optimal strategy. The authors show in the paper how such scenarios can be modeled in a principled way, so that the recommendation is not based on an idealized “empty city”, but on competition for a limited resource.
How big is the gain in practice: Seattle as a testbed
To test the potential, the researchers ran simulations using real traffic data from the Seattle area. In the most congested scenarios, the approach that takes parking probability into account achieved time savings of up to 66 percent compared to strategies that ignore parking and simply guide the driver to the nearest parking facility or leave them to “wait for luck” in the nearest garage. Translated into the experience of a single driver, this can mean reducing total travel time by up to about 35 minutes in the worst conditions, when the difference between “park closest” and “park smarter” increases dramatically.
At the same time, the authors warn that even the best “parking-aware” strategy can take noticeably longer than the optimistic “directly to the destination” estimate shown by classic navigation. In their experiments, total door-to-door time can be up to 123 percent longer than an estimate that ignores parking. This is an important message for users and cities: the problem is not only that the trip is “sometimes a bit longer”, but that the standard estimate is often structurally biased against the real experience.
Where availability data comes from: sensors, gates, or — people
The biggest obstacle to scaling such a system is not the math itself, but the data: how likely is it that a particular parking facility is free at a particular moment? Some parking facilities have gates, entry/exit counters, or magnetic detectors that can provide a relatively reliable picture of occupancy. But such infrastructure is not universal, and its rollout is expensive and slow.
That is why the researchers also consider crowdsourcing, i.e., collecting observations from user behavior. The idea is that “signals” about occupancy can be obtained from multiple channels: a user in the app can mark “no parking”, vehicles that enter a garage and quickly exit without parking can signal failure, and the number of cars circling around the block can indicate high occupancy. In perspective, autonomous vehicles could automatically record free spaces they spot while passing by, creating a continuous stream of observations without additional burden on drivers.
In their error analysis of availability estimation, the authors state that when relying on stochastic observations one can obtain a mean absolute error of around 7 percent, and with more frequent observations it can be reduced below 2 percent. This suggests that even without full sensor coverage the system can be accurate enough to provide the user with a more realistic arrival estimate and a more meaningful parking recommendation.
What such navigation would change in everyday city life
The most direct benefit is predictability. Instead of the driver getting an estimate to the address and only then realizing that an uncertain 10 or 20 minutes awaits, they would get an overall arrival estimate that includes parking and walking. This changes day planning, but also the way we compare transport options. If an app shows a realistic “door-to-door time” for a car, and for public transport the actual time with transfers and walking, the user has a clearer basis for a decision.
The second benefit is operational: reducing cruising can relieve key city streets and intersections, especially around the most burdened zones. Although this MIT paper primarily measures time savings, the authors point out that a potential next step is estimating emission reductions. The logic is simple: fewer “aimless” kilometers means less fuel and fewer exhaust gases, and in peak periods fewer secondary slowdowns that cruising can trigger.
The third benefit concerns parking policy. Cities that manage parking through pricing, time limits, and digital rule maps often try to achieve a balance — enough turnover so spaces are not constantly occupied, but also enough predictability so the system is acceptable to users. Information about probability of success, if shown transparently, can be an additional tool: the driver can consciously choose slightly farther parking with a higher chance of success, instead of “everyone going to the same place” and thereby creating a bottleneck.
Limitations and open questions: privacy, behavior, and integration with existing apps
The authors emphasize that their work is not yet a finished product ready for mass deployment, but a feasibility demonstration. To introduce the system into a real city, a series of practical questions must be resolved. One is privacy: crowdsourcing and tracking entries/exits from parking facilities can be sensitive if not carried out with clear anonymization policies and minimal data collection. The second is user behavior: as soon as recommendations become popular, they can change traffic flows and “overload” the parking facilities the algorithm often suggests, so the model must have stabilization and adaptation mechanisms.
The third is integration. Today’s navigation systems already combine road data, traffic conditions, and congestion predictions, but parking is often “outside the equation” or displayed as a list of parking facilities without a reliable availability estimate. Introducing parking probability would require standardized data flows between cities, private garage operators, and the platforms users actually use. In that sense, it is also useful that the authors published accompanying code alongside the paper, which facilitates checks, replications, and upgrades of the approach.
Broader context: a small shift in information, a big shift in outcomes
In transport systems, changes are slow because they involve infrastructure, habits, and policy. But research like this often targets “small levers”: if we give the user more accurate information at the right moment, the behavior of thousands of people can change without building new roads. The MIT model starts from a simple idea — that parking is part of the trip — and shows that this can be mathematically incorporated into the route in a way that reduces expected arrival time and, indirectly, traffic pressure in the center.
If further studies are confirmed in real time and at the scale of entire cities, parking-aware navigation could become a standard that changes expectations: the estimate of a car trip would no longer end at “arrival at the address”, but at actual arrival at the door. For the driver that means less uncertainty, for the city potentially less cruising and gridlock, and for all other traffic participants — a more predictable and cleaner urban space.
Sources:- arXiv – preprint of the paper “Probability-Aware Parking Selection” (C. Hickert, S. Li, Z. He, C. Wu) with a description of the model, simulations, and results ( link )- arXiv (PDF) – details on parking availability estimation errors and ranges of time savings in simulations ( link )- GitHub – code repository associated with the paper “Probability-Aware Parking Selection” ( link )- UCLA / Donald Shoup – review “Cruising for parking” with analysis of earlier studies on search time and the share of traffic that cruises ( link )- Springer Nature (Transportation) – open paper on predicting cruising time and estimating emissions from parking search in dense urban zones ( link )
Find accommodation nearby
Creation time: 4 hours ago