PaperlayerView exact revision on arXiv

Source-linked AI summary

RoboSense: Leveraging Robotaxi Fleets as Drive-by Sensors for Urban Traffic Monitoring

Yilin Wang, Yiheng Feng

arXiv:2609.06813v1cs.ROcs.SYeess.SY

TL;DR

Fixed sensors are spatially limited, while probe vehicles provide insufficient or uncontrolled coverage for network-level monitoring. This paper uses a cell-based, rolling-horizon MILP to coordinate robotaxi routing for travel time and coverage. Monitoring-aware routing improves coverage, and appropriate objective weights can also improve average speed.

  • Problem

    Fixed sensors have limited spatial coverage, while probe vehicles need sufficient market penetration and follow uncontrolled personal travel patterns, complicating spatiotemporal traffic monitoring.

    Source: §1 · p. 1

  • Method

    A cell-based, rolling-horizon MILP coordinates robotaxi routes to minimize travel time while maximizing spatiotemporal network coverage.

    Source: §6 · p. 25

  • Results

    Monitoring-aware routing improves spatiotemporal network coverage, and appropriate objective weights can also increase robotaxi average speed.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Coordinated robotaxi fleets have potential to serve as drive-by sensors for urban traffic monitoring and may incentivize operator participation through mobility benefits.

    Source: §1 · p. 3

    The framework uses simplified settings; future work includes more realistic perception uncertainty, passenger pickup and drop-off operations, and integration with fixed sensors.

    Source: §6 · p. 25

Abstract

from arXiv · show

Urban traffic monitoring plays a critical role in safety analysis, congestion management, and incident response. The growing deployment of robotaxis creates a new opportunity for network-level traffic monitoring. Although robotaxis are primarily designed to serve passengers, they can also be leveraged as drive-by sensors to collect traffic data. Compared to conventional infrastructure sensors or probe vehicles, a fleet of robotaxis forms a cooperative perception environment, which can collectively gather spatially and temporally continuous traffic information. This paper proposes a novel dynamic robotaxi routing framework that explicitly incorporates traffic monitoring tasks as an objective. The framework introduces: (1) a cell-based network representation that aligns with sensing capabilities of robotaxis; (2) a cell-level monitoring metric to quantify spatiotemporal robotaxi coverage; and (3) a mixed-integer linear programming (MILP) formulation that jointly minimizes time-dependent travel time and maximizes traffic monitoring performance. A 5 by 5 urban grid network is built in SUMO to evaluate the framework under three robotaxi market penetration rates (2%, 5%, and 10%) with a range of objective weight combinations. Results show that incorporating spatiotemporal network coverage in the objective function can effectively improve the traffic monitoring performance. Interestingly, with appropriate weights between the two objectives, monitoring performance and robotaxi average speed can be improved simultaneously. This suggests better network monitoring leads to more accurate traffic state prediction and improved mobility. This win-win situation could incentivize robotaxi operators to contribute their vehicles as drive-by sensors for traffic monitoring.

1. Introduction

Robotaxis could address the spatial limits of fixed sensors and the low coverage and uncontrollable travel patterns of probe vehicles. This paper coordinates robotaxi routing for traffic monitoring and reports that monitoring coverage can improve without sacrificing speed under appropriate objective weights.

1. Introduction

  • Fixed sensors provide temporally continuous but spatially limited data, while probe vehicles require sufficient market penetration and follow trips they cannot control.Probe vehicles also report their own trajectories rather than surrounding traffic.Source: §1 · p. 1
  • Robotaxi fleets offer cooperative perception and centralized routing, enabling coordinated data collection while meeting passenger travel-time requirements.Unlike probe vehicles, robotaxis can perceive surrounding traffic and share perception data in real time.Source: §1 · pp. 1–2
  • Routing couples mobility and monitoring because vehicle paths determine which network areas are observed and when.The example contrasts redundant shortest-path observations with an alternative route that observes an otherwise unobserved corridor.Source: §1 · p. 2
  • A higher-resolution cell-based network is needed because robotaxi sensing ranges are shorter than typical urban road links, leaving only portions of links observed.Reported sensing ranges are about 60–120 m.Source: §1 · p. 2
  • Monitoring needs a network-level metric based on individual robotaxi observations that captures the spatial and temporal distribution of coverage.The metric must reflect monitoring quality rather than travel efficiency.Source: §1 · p. 2
  • Dynamic routing must adapt to changing travel times and network observability across time, vehicles, and network states.Combining this high-dimensional decision space with multiple objectives increases problem complexity.Source: §1 · p. 2
  • Monitoring-aware routing substantially improves spatiotemporal network coverage, and appropriate objective weights can also increase robotaxi average speed.The framework treats monitoring and travel time as joint routing objectives.Source: §1 · p. 3
  • The contributions are an integrated routing-and-monitoring framework, a cell-level observability metric, and simulation evaluation across robotaxi market-penetration scenarios.The framework jointly represents vehicle mobility and network observability.Source: §1 · p. 3

2. Literature Review

The review covers fixed sensors, probe vehicles, and cooperative perception, then surveys vehicle-routing models and their objectives. It identifies limited work on dynamically coordinating robotaxi fleets for traffic monitoring.

2.1. Traffic Monitoring

  • Fixed sensors provide temporally continuous point observations, prompting methods to infer spatial traffic states from sparse measurements.Prior approaches include statistical inference, kinematic-wave models, learning-based estimation, and camera-derived trajectories.Source: §2.1 · pp. 3–4
  • Probe-vehicle trajectories support state estimation and reconstruction, but low penetration and passive, demand-driven travel constrain their monitoring coverage.The review contrasts these passive observations with coordinated robotaxi sensing.Source: §2.1 · p. 4
  • Robotaxi fleets can share perception data at lower market penetration than probe vehicles, but clustered and overlapping observations require new estimation methods and coordinated fleet distribution.Few studies actively manage fleet mobility to reduce redundant observations.Source: §2.1 · p. 4

2.2. Vehicle Routing Models

  • Vehicle-routing models commonly minimize travel cost or congestion, with other work adding objectives such as energy use, fairness, or sensing benefit.The reviewed sensing-oriented routing work treats traffic conditions and sensing structure as exogenous inputs.Source: §2.2 · p. 5
  • The review identifies few studies that integrate traffic monitoring into dynamic routing, particularly for centrally coordinated robotaxis with cooperative perception.This paper presents its work as among the earliest efforts to address that gap.Source: §2.2 · p. 5

3.1. Framework Overview

The framework connects SUMO vehicle movement, cell-based observation, CTM traffic-state estimation, and MILP routing in a rolling-horizon loop. It assumes a central controller coordinates robotaxis and that their routing does not alter background-vehicle routing or travel times.

3.1. Framework Overview

  • SUMO robotaxi locations and sensing ranges determine observed cells, using cell lengths matched to typical robotaxi detection ranges.An observed cell is treated as providing complete traffic data, including average speed and density or vehicle counts.Source: §3.1 · p. 5
  • The CTM estimates and predicts cell traffic states, from which the framework derives time-dependent travel costs for routing.The prediction horizon matches the dynamic vehicle-routing planning horizon.Source: §3.1 · p. 6
  • A rolling-horizon MILP jointly minimizes robotaxi travel time and maximizes observed cells, then sends optimized routes back to SUMO for execution.The routing model uses network structure, robotaxi locations, origin-destination information, and predicted travel times.Source: §3.1 · p. 6
  • The framework assumes robotaxi routing does not change background traffic routes or travel times, and a central controller has real-time fleet information and route compliance.Intersection travel time is omitted, although queue and signal delays are counted.Source: §3.1 · p. 6

3.2. Cell Transmission Model for Urban Network

The urban-network CTM represents links and intersections with specialized cells and models traffic flow between them. At diverging cells, the framework relaxes FIFO to reflect lane changes in the SUMO setting.

3.2. Cell Transmission Model for Urban Network

  • The CTM is a first-order discrete approximation of the LWR model, representing roads as consecutive homogeneous sections with a triangular or trapezoidal fundamental diagram.Its network extension represents urban roads with signalized intersections.Source: §3.2 · p. 6
  • Six cell types represent the urban network, with entrance flows merging through link cells before diverging toward downstream intersection cells.The middle link cells have twice the capacity of other cells.Source: §3.2 · pp. 6–7
  • The model relaxes FIFO at diverging cells so flows can enter downstream cells independently, reflecting lane changes that balance queues in SUMO.Diverging ratios use historical turning flows and downstream receiving capacities, and vary over time.Source: §3.2 · pp. 7–8

3.3. Dynamic Vehicle Routing Model Formulation

The MILP jointly minimizes robotaxi travel time and maximizes spatiotemporal cell coverage, using time-dependent travel costs and constraints that encode feasible routes and cell occupation. Its formulation also prunes infeasible time-dependent arcs to reduce complexity.

3.3. Dynamic Vehicle Routing Model Formulation

  • The model represents CTM cells as nodes in a directed graph and discretizes time over a planning horizon long enough for robotaxis to reach their destinations.Its variables encode selected route cells and entry times, downstream-cell and departure-time connections, and cell occupation for coverage.Source: §3.3 · p. 8

3.3.1. Objective function

  • The objective balances total robotaxi travel time against normalized spatiotemporal network coverage using weights α1 and α2.Coverage sums cell occupation over time and is normalized by the number of time steps and cells.Source: §3.3.1 · pp. 8–9

3.3.2. Variable 𝑥- 𝑧connection constraints

  • Constraints link selected cells to downstream cells and departure times, allowing only connected downstream cells in the network topology.The variables encode one combination of departure time and downstream cell when a robotaxi enters a cell.Source: §3.3.2 · p. 9

3.3.3. Network flow constraints

  • Flow constraints enforce time-continuous feasible paths from each robotaxi’s origin to destination, with no simultaneous entry into multiple cells or repeat visits.Intersection travel time is not modeled separately; signal waiting time is included in cell travel time.Source: §3.3.3 · p. 9

3.3.4. Vehicle-level network occupation constraints

  • Vehicle-level occupation ω is tied to route selections: it is zero when a robotaxi never visits a cell and tracks occupation during its visit.The occupation constraints are linearized with the big-M method.Source: §3.3.4 · p. 9

3.3.5. Cell-Level network occupation constraints (𝜔- 𝑦relation)

  • Cell-level occupation y aggregates robotaxi occupation so simultaneous occupancy by multiple robotaxis can be handled; the study assumes perfect detection of vehicles in an occupied cell.The authors note that the model could be extended to detection models accounting for accuracy and uncertainty.Source: §3.3.5 · p. 11
  • Time-dependent cell travel time is calculated from CTM-predicted traffic states, with cell speed derived from a trapezoidal fundamental diagram.The traverse time is the number of time steps needed for accumulated distance to reach the cell length.Source: §3.3.5 · p. 11

3.3.6. Complexity Analysis

  • The formulation prunes infeasible time-dependent arcs in advance; with sparse connections, z-variable dimension and total constraints scale as Θ(|A||E||T|) and Θ(|A||I||T|), respectively.Here |E| is the number of directed cell-to-cell connections; the model describes the z-variable dimension as linear in robotaxis, connections, and planning horizon.Source: §3.3.6 · p. 12

4. Case Study

The case study first uses a toy network to illustrate routing under different objective weights, then sets up a 5 × 5 urban network for comprehensive evaluation. In the toy example, emphasizing coverage spreads routes across the network but increases total travel time.

4. Case Study

  • The case study uses a toy network to demonstrate routing strategies, followed by a 5 × 5 urban network for comprehensive evaluation.The supplied passage does not specify further details of the larger network setup.Source: §4 · p. 12

4.1. Toy Network Analysis

  • With four vehicles on the 26-cell toy network, coverage-focused weights produce different routes and greater spatiotemporal coverage than travel-time-only routing, at the cost of increased total travel time.Under travel-time-only weights, all four vehicles choose the same minimum-time route, creating redundant observations; Figure 7 presents route choices and coverage under the two weight settings.Source: §4.1 · p. 12, Figure 7 · p. 13, Figure 8 · p. 13

4.2. Grid Network Analysis

The grid-network evaluation tests dynamic routing under three robotaxi penetration rates, with monitoring coverage weighted alongside travel time. Results show coverage generally increases with its objective weight, while the routing strategy and sensing power depend on fleet size and mobility.

4.2.1.1 SUMO configuration

  • The SUMO evaluation uses a 5 × 5 grid with 760 CTM cells, 16 OD pairs, and fixed-time signals; robotaxi routes are updated every 100 s.The network has 40 bidirectional two-lane links and 25 intersections; each internal link has seven 80 m cells.Source: §4.2.1.1 · pp. 13–15

4.2.1.2 Robotaxi De-routing Budget

  • A de-routing budget limits each robotaxi to at most two extra links, and disables detours when the remaining route exceeds six links or the vehicle has traveled over 12 links.These rules constrain repeated subtours and reduce the feasible routing choices.Source: §4.2.1.2 · p. 15

4.2.2. CTM Calibration and Evaluation

  • The calibrated triangular CTM fundamental diagram uses q_max = 1,920 veh/h and k1 = k2 = 33.33 veh/km, with an entry-link MAE of 0.38 veh/(cell·step).The passage attributes the main error source to fixed CTM turning ratios that may not match signal-cycle turning ratios; Figures 12 and 13 show calibration and a link-level comparison.Source: §4.2.2 · pp. 15–16, Figure 12 · p. 16, Figure 13 · p. 16

4.2.3. Solving Time Analysis

  • At 10% MPR, routing about 142 vehicles per iteration takes an average of 160 seconds, and average solving time remains below 200 seconds.Computation time increases with the number of robotaxis.Source: §4.2.3 · p. 16, Figure 14 · p. 17

4.2.4. Result and Discussion

  • Increasing α2 raises spatiotemporal coverage across MPRs, but higher monitoring weights do not always improve sensing power.The study evaluates network and vehicle coverage across MPRs and objective-weight combinations.Source: §4.2.4 · pp. 16–18, Table 3 · p. 18

4.2.4.1 Robotaxi Mobility and Spatiotemporal Coverage

  • Average travel time converges to about 1,000–1,100 s as α2 increases, while coverage reaches its ceiling at α2 = 10,000, 3,000, and 1,500 for 2%, 5%, and 10% MPR, respectively.Figure 15 plots average travel time and network coverage; the authors attribute the lower ceiling thresholds at higher MPRs to greater routing flexibility.Source: Figure 15 · p. 17, §4.2.4.1 · pp. 18–19
  • At 10% MPR without a de-routing budget, CTM MAE fell from 0.99 to 0.74 and average travel time fell from 763.94 s to 723.84 s as α2 rose from 10 to 1,500.Average travel distance stayed around 2,940 m; at still higher α2, average travel time began increasing as the coverage objective dominated.Source: §4.2.4.1 · p. 19

4.2.4.2 Sensing Power Evaluation

  • Sensing power peaks at α2 = 10,000, 3,000, and 1,000 for 2%, 5%, and 10% MPR, respectively, rather than at the highest α2.After de-routing budgets are used, higher α2 can reduce speed while segment popularity remains similar, lowering sensing power.Source: §4.2.4.2 · pp. 19–20, Figure 16 · p. 20

4.3. Vehicle Number Weighted Monitoring Objective

The vehicle-number-weighted objective directs sensing toward cells with more vehicles while keeping the model as an MILP. It increases observed-vehicle coverage, but can route robotaxis through more congested cells and reduce average speed.

4.3. Vehicle Number Weighted Monitoring Objective

  • The revised objective weights each cell’s coverage by its vehicle count, which the CTM supplies as a constant coefficient, so the formulation remains an MILP.The experiments revise α2 values to account for the modified coverage term’s magnitude.Source: §4.3 · p. 22
  • Vehicle weighting routes robotaxis toward more congested cells, raising coverage but reducing average speed, especially at 2% MPR.At higher MPRs, assigned routes are more diverse and include less-congested links; travel times converge between 1000s and 1200s.Source: §4.3 · pp. 22–23, Figure 21 · p. 23
  • The vehicle-weighted objective yields higher average observed-vehicle counts across robotaxi MPRs, with counts increasing as α2 rises.Under the cell-based objective, average observed counts remain mostly constant across α2 values.Source: §4.3 · p. 23, Figure 22 · p. 24
  • Weighting shifts sensing effort from cell-time coverage toward vehicle-time observations, allowing task-specific emphasis based on traffic conditions and policy preferences.Source: §4.3 · p. 23

5. Implications of Real-world Deployment

Robotaxis already carry perception and communication hardware and spend some time without passengers, creating an opportunity to use them for coordinated monitoring. The paper also describes passenger incentives and conditional cases where monitoring and mobility objectives improve together.

5. Implications of Real-world Deployment

  • Robotaxis’ existing perception and communication hardware, together with time spent without passengers, make them an underused potential sensing layer.Source: §5 · p. 24
  • Passenger incentives could encourage acceptance of slightly longer trips in exchange for improved traffic monitoring.The passage cites incentive-based routing studies and Metropia as a real-world implementation example.Source: §5 · p. 25
  • Under some conditions, improved network coverage coincides with lower robotaxi travel time, which the paper links to more accurate travel-time prediction and route selection.Source: §5 · p. 25

6. Conclusion and Future Work

The paper presents a robotaxi routing-for-monitoring framework and evaluates it on a 5 × 5 SUMO grid. Results indicate that objective weights can alter routes and sometimes reduce travel time, while excessive detouring can degrade sensing power; future work targets more realistic deployment conditions.

6. Conclusion and Future Work

  • The framework combines microscopic simulation, traffic-state estimation, and dynamic robotaxi routing, optimizing travel time and spatiotemporal network coverage through a rolling-horizon MILP.The case study uses a 5 × 5 grid with symmetric OD demand and fixed-time signals, testing 2%, 5%, and 10% MPRs.Source: §6 · p. 25
  • Appropriate objective weights can reduce robotaxi average travel time, while detouring that only extends distance can lower fleet speed and sensing power.The authors conclude deployment must balance coverage uniformity, fleet speed, and travel-time penalty rather than only maximize coverage.Source: §6 · p. 25
  • The current work’s simplified settings motivate future extensions for perception uncertainty, passenger pick-up and drop-off operations, and integration with fixed-location sensors.Source: §6 · p. 25

8. CRediT authorship contribution statements

The contribution statement assigns writing, visualization, validation, methodology, investigation, formal analysis, data curation, and conceptualization to Yilin Wang, and writing, project administration, methodology, investigation, funding acquisition, and conceptualization to Yiheng Feng.

8. CRediT authorship contribution statements

  • Yilin Wang is credited with visualization, validation, formal analysis, and data curation; Yiheng Feng is credited with project administration and funding acquisition.Both are credited with writing, methodology, investigation, and conceptualization.Source: §8 · p. 25
Loading 2609.06813v1…