Operations of Shared Autonomous Vehicle Fleet for Austin, Texas, Market

The paper explores how shared autonomous vehicles (SAVs) may operate and affect urban travel, focusing on a realistic network and demand context in Austin, Texas. As multiple automakers and technology firms advance toward highly automated driving, SAVs—driverless, on-demand vehicles—could merge the advantages of car sharing and taxis by traveling unoccupied to customers and being centrally managed for system-optimal operations. Compared to traditional car-sharing, SAVs eliminate the need to keep vehicles at destination and can preemptively relocate; compared to human-driven taxis or transportation network companies (TNCs), automation can reduce labor costs and enable fully centralized routing and relocation for improved system performance. Given these potential advantages, SAVs could capture significant shares of personal travel and influence private vehicle ownership. The study aims to quantify operational performance, replacement rates of conventional vehicles, wait times, empty VMT, and environmental implications at low market penetration using a time-varying speed network model.

The authors situate their work within prior SAV and taxi literature. Burns et al. estimated large cost reductions for autonomous taxis in Manhattan by replacing driver labor, though with optimistic automation cost assumptions. Earlier simulations by Fagnant and Kockelman suggested SAV fleets could reduce taxi fares while yielding operator returns and examined relocation strategies in idealized grids. Kornhauser et al. coined the term aTaxis and studied areawide concepts; Pavone et al. developed load balancing for mobility-on-demand systems. Comparisons are drawn with TNCs (Uber/Lyft) that are centrally coordinated but still rely on driver decisions. This paper advances prior work by using an actual metropolitan network (Austin), with time-dependent link speeds (from MATSIM) and empirically grounded, time-of-day demand distributions, to assess SAV operations, relocation, and environmental outcomes.

Study area and demand: The Austin regional network and trip tables from CAMPO (six counties; 2,258 TAZs; 13,594 nodes; 32,272 links) were used. A 12-mi by 24-mi central geofence was defined due to high trip densities. A 24-hour weekday demand profile was created using 4.5 million trips (including commercial), with MATSIM run on a 5% sample (vehicles represent 20 cars) to produce hourly link travel times. Departure time distributions were derived from 2006 Seattle household travel diaries for smooth, realistic peaks. A 100,000-trip subset was randomly drawn; 57,161 travelers (1.3% of internal trips) within the geofence were assumed to use SAVs; trips crossing the geofence boundary used other modes. Simulation framework: A custom C++ program simulated SAV operations with 5-minute assignment intervals and time-dependent link speeds. Four submodules: (1) SAV location and trip assignment using a backward-modified Dijkstra search to find the nearest available SAV, with an initial 5-minute max path time (expanding to 10 minutes for those on a wait list, prioritizing longest-waiting travelers); time-dependent shortest paths guide pickups and deliveries; 1-minute pickup and drop-off service times are applied. (2) SAV fleet generation via a seed day: when a traveler waits 10 minutes without an available SAV within 10 minutes, a new SAV is instantiated at that location and kept in the fleet; the measured day starts with the seed-day end-of-day vehicle locations and no new vehicles added. (3) SAV movement: vehicles advance along paths with progress saved across intervals; they can serve multiple actions in a 5-minute window if time allows. (4) SAV relocation: a 2-mi by 2-mi block-balancing heuristic minimizes imbalances between SAV supply and expected demand (waiting trips plus expected next 5-minute arrivals based on 1-hour bin averages). Block balance = SAVs_block − (SAVs_block/SAVs_total)*(demand_block/demand_total). Blocks with balance < −5 pull from neighbors with highest positive balances; > +5 push to neighbors with lowest balances, requiring an inter-block balance difference > 1. SAVs chosen for relocation minimize travel time to destination block centers; relocation is reassessed every 5 minutes, with stopping rules to avoid over-concentration at centers. Alternative finer-grained relocation heuristics from prior work were not used due to lower effectiveness in the realistic network context. Application: One seed day determined fleet size; a subsequent day with the same starting trip population evaluated performance: fleet size, trips served per SAV, wait times, speeds, VMT including empty travel, and environmental metrics. EV charging feasibility was explored using stationary interval distributions and daily miles per SAV. Life-cycle emissions were computed using Chester and Horvath factors, comparing an SAV passenger-car fleet to the current U.S. light-duty fleet mix.

• Fleet sizing and service: Approximately 1,977 SAVs served the geofenced demand sample. Each SAV averaged 28.5 person trips/day.
• Replacement rate: Using 3.02 person trips/day per licensed driver and 0.99 licensed drivers per household vehicle, each SAV could replace about 9.34 conventional vehicles within the geofence, acknowledging this is an upper-bound under pure substitution and ignoring external trips.
• Wait times: Average wait time ≈ 1 minute; 94.3% waited < 5 minutes; 98.8% < 10 minutes; 0.10% waited 15–29 minutes. Peak hour (5–6 p.m.) average wait reached 3.85 minutes. Actual call times between 5-minute intervals would add on average 2.5 minutes to user waits.
• Empty VMT: Approximately 8.0% additional VMT arose from unoccupied SAV movements (pickups and relocations).
• Speeds: Distance-weighted speeds averaged 43.6 mph over 24 hours; time-weighted average speed (VMT/VHT) was 26.1 mph. About 19.4% of VMT occurred at ≤20 mph; 41.4% at >50 mph.
• Comparison to NYC taxis: NYC yellow taxis average 36 trips/day, 2.6-mi average trip, ~70,000 mi/year with ~51.5% empty VMT, versus the simulated 8.0% empty VMT in Austin. Austin covers a much larger service area with longer average trips (5.2 mi).
• EV feasibility: SAVs averaged 174 miles/day; each had on average 2.91 stationary intervals ≥1 hour and 0.80 intervals of 30–59.9 minutes per day, potentially usable for charging. Typical BEV ranges (60–100 mi) may be insufficient without fast charging or higher-range vehicles (e.g., Tesla Model S 208–265 mi), making PHEVs or fast charging important in early deployments.
• Emissions (per SAV introduced, life-cycle comparison to U.S. fleet average passenger vehicles): Energy use −14%; GHG −7.6%; SO₂ −20%; CO −32%; NOₓ −18%; VOC −47%; PM₁₀ −7.6%. Cold starts reduced by ~85% due to higher vehicle utilization and fewer long rests.
• Relocation heuristic effectiveness: The 2-mi block balancing notably reduced wait times in prior work and is well-suited to Austin’s centralized demand; thresholds used here were ±5 vehicles per block and >1 inter-block balance difference to justify moves.
• Parking implications: If each SAV replaces ~9 conventional vehicles, on the order of 8 parking spaces per SAV could be freed, enabling alternative land uses.

The study demonstrates that in a realistic, time-dependent network and at low market penetration, a centrally managed SAV fleet can achieve short user wait times and high replacement rates of private vehicles with modest increases in empty VMT. The relocation strategy balances traveler wait reduction with empty miles, achieving an 8% increase in VMT while maintaining average waits near 1 minute, indicating effective supply–demand matching within the dense geofence. The results indicate that SAV efficiency improves with higher demand density: more trips per vehicle, lower wait times, and lower empty VMT share. Comparisons with an idealized grid model highlight that while idealized settings can suggest even higher replacement rates and lower waits, realistic network circuity, heterogeneous demand, and variable speeds constrain performance; nevertheless, the Austin scenario still performs strongly. Environmental analysis suggests that despite extra empty miles, fleet downsizing to smaller passenger cars, significant reductions in cold starts, and higher utilization can reduce life-cycle energy use and pollutant emissions. Operationally, the daily mileage and idle-period distributions suggest that electrified SAV fleets are feasible with appropriate charging strategies, especially using higher-range BEVs, PHEVs, or fast-charging infrastructure. The parking and land-use benefits further underscore SAVs' broader urban impacts. Overall, the findings support the hypothesis that SAVs can efficiently serve intra-urban trips and materially reduce private vehicle dependence and environmental burdens, provided complementary policies and infrastructure (e.g., charging, curb management) are in place.

Simulations for Austin’s 12×24-mile core suggest that a fleet of roughly 1,977 SAVs can serve a low-penetration demand sample with high service quality: average 1-minute waits, about 28.5 trips per SAV per day, and an estimated replacement rate of roughly 9.3 conventional vehicles per SAV under pure substitution within the geofence. The operational trade-off includes about 8% additional VMT from unoccupied movements. Nonetheless, life-cycle energy and emissions are expected to decline due to fewer cold starts and a lighter passenger-car fleet. Parking demand could fall substantially, freeing urban land for higher-value uses. Electrification appears promising given daily mileage and idle windows, especially with fast charging or longer-range/PHEV options. Future research should integrate mode choice and vehicle ownership responses, examine multi-day stochastic demand, optimize relocation via explicit objective functions, assess multi-hub geographies with separated demand centers, and evaluate dynamic ride-sharing and pricing strategies that further reduce empty miles and improve equity and efficiency.

• Geofenced analysis: Only trips fully inside the 12×24-mile area are served by SAVs; external trips rely on other modes, potentially biasing replacement-rate estimates upward.
• Demand sample and penetration: Results reflect 1.3% of internal trips and may not scale linearly with higher market penetration.
• Time discretization: Vehicle assignment and relocation operate on 5-minute intervals; travelers are assumed to request at these boundaries in the model (real-world off-cycle requests add ~2.5 minutes expected wait).
• Network speeds: Hourly link travel times come from a 5% MATSIM simulation (vehicles represent 20), which reduces fidelity but captures congestion patterns; exact speeds and congestion could differ under full-scale simulation.
• Departure time data: Seattle 2006 household diaries were used to shape Austin’s departure distribution due to issues in the local survey, introducing cross-city transfer assumptions.
• Fleet generation: A single seed day establishes fleet size; only one subsequent day is analyzed, not a full distribution over many days, though prior work suggests stability.
• Simplifications: Pure substitution of household vehicle trips is assumed for estimating vehicle replacement; taxi/TNC substitution share is small but non-zero. Mode choice, vehicle ownership dynamics, and behavioral adaptation are not endogenously modeled. Advance reservations, no-show behavior, multi-passenger ride-sharing, and pricing strategies are not represented.
• Relocation heuristic: Uses thresholds and a block-based approach; not an optimization of a formal cost function. Effectiveness may vary in polycentric regions separated by low-demand areas.
• Vehicle heterogeneity: SAVs modeled as passenger cars; impacts may differ with mixed vehicle sizes or shared-ride policies.
• Electrification analysis is indicative: Charging logistics, station siting, queuing, and battery degradation are not explicitly modeled.