Autonomous Taxi Fleet Optimization Boosts Profit and Efficiency
A groundbreaking study from Guangdong University of Technology introduces a new dynamic scheduling model for autonomous electric taxi fleets that significantly enhances long-term operational profitability by integrating future revenue predictions into real-time decision-making. Led by Zeng Weiliang and Han Yu from the School of Automation, along with Fu Hui from the School of Electromechanical Engineering, the research leverages reinforcement learning to optimize both routing and charging strategies, marking a major advancement in smart urban mobility systems.
The study, published in Industrial Engineering Journal, addresses a critical limitation in existing taxi dispatch models—namely, their focus on immediate cost optimization at the expense of long-term operational value. Traditional systems often rely on greedy matching algorithms that assign nearby vehicles to incoming ride requests without considering how such decisions affect future availability, positioning, or energy costs. While these methods are computationally efficient, they frequently lead to suboptimal fleet distribution, increased passenger wait times, and inefficient energy use, especially in large-scale autonomous fleets operating around the clock.
To overcome these shortcomings, the team developed a novel framework that evaluates not only the immediate reward of each dispatch decision but also its potential impact on future earnings. This forward-looking approach is made possible through the use of a state-value function approximated via deep neural networks, which estimates the long-term cumulative return associated with a vehicle’s spatiotemporal state after executing a given action. By incorporating this predicted future value into the current dispatch decision, the model enables more strategic vehicle repositioning, charging, and passenger matching.
The core innovation lies in the integration of vehicle-to-grid (V2G) technology and time-of-use electricity pricing into the dispatch logic. Unlike conventional models that treat charging as a necessary but passive activity, this system actively uses charging stations as microgrids where idle autonomous taxis can either charge during off-peak hours or discharge back to the grid during peak demand periods, thereby generating additional revenue. This bidirectional energy flow transforms electric vehicles from mere transportation units into mobile energy assets that contribute to grid stability while reducing fleet operating costs.
The model operates within a fixed time window—set at five minutes in the simulation—during which the system evaluates all available actions for each “dispatchable” vehicle. These actions include idling, relocating to a new area, charging, discharging, or accepting a passenger request. Each action is assigned a composite value that combines its immediate financial return with the discounted future value of the resulting vehicle state. The global dispatch solution is then determined by solving a weighted bipartite matching problem that maximizes total fleet value while respecting constraints such as one vehicle per trip and one trip per vehicle.
To ensure computational efficiency and scalability, the researchers employed a two-stage solution process. First, a greedy heuristic identifies a high-quality initial matching based on action-value rankings, prioritizing passenger pickups when values are comparable. This initial solution is then refined using an integer programming solver, allowing the system to handle large-scale urban environments with hundreds of vehicles and thousands of nodes. The entire process is repeated at each time step, enabling continuous adaptation to evolving traffic conditions and demand patterns.
Training the state-value function relies on a deep reinforcement learning architecture enhanced with double neural networks and experience replay, techniques commonly used in deep Q-learning. An evaluation network estimates the current value of vehicle states, while a separate target network provides stable training signals by computing bootstrapped future returns. Experience tuples—comprising state, action, reward, and next state—are stored in a replay buffer and sampled in mini-batches to update the evaluation network, improving data efficiency and reducing correlation between consecutive updates. The exploration-exploitation trade-off is managed through an ε-greedy policy, where the system initially explores random actions to gather diverse experiences before gradually shifting toward value-driven decisions as the model converges.
The experimental validation was conducted using a simulated road network extracted from Shenzhen, covering a 6 km × 6 km urban area with 2,876 roads and 1,712 intersections. The test fleet consisted of 40 autonomous electric vehicles, each equipped with a 200 kWh battery, capable of charging and discharging at 150 kW. Ride requests were generated according to a Poisson process calibrated to historical taxi demand data, with origin probabilities concentrated toward the city center and destinations randomly assigned. Passenger wait tolerance was modeled as a uniform random variable between 10 and 30 minutes, reflecting realistic user expectations.
Commercial time-of-use electricity rates in Shenzhen were incorporated into the simulation, featuring distinct peak, shoulder, and off-peak periods. During peak hours (10:00–15:00 and 18:00–21:00), electricity prices reached 1.38 yuan/kWh, while off-peak rates dropped to 0.28 yuan/kWh between 23:00 and 07:00. This significant price differential creates strong economic incentives for intelligent charging and discharging strategies, which the proposed model exploits to minimize net energy costs.
Performance was benchmarked against three established approaches: First-Come-First-Served (FCFS), which assigns the nearest available vehicle to each request; Optimal Benefit (OB), which maximizes immediate revenue at each time step; and a standard Deep Q-Network (DQN) model adapted for multi-vehicle dispatch. All algorithms were tested over a simulated 100-day operational period, with results averaged across 10 independent test sets to ensure statistical robustness.
The findings revealed substantial advantages for the proposed method. Over the 100-day simulation, the new model achieved a 25% increase in total operational profit compared to the best-performing baseline. This gain stems from a combination of higher service rates, reduced energy expenses, and improved fleet utilization. Specifically, the passenger response rate—the proportion of ride requests successfully fulfilled—rose by 4 percentage points, reaching over 73% compared to 69% for the DQN and below 60% for FCFS and OB. Simultaneously, average passenger waiting time decreased by 20%, from over 20 minutes under FCFS to just under 8 minutes, enhancing user satisfaction and competitiveness in the mobility market.
Energy cost reduction was particularly striking. Despite higher overall electricity consumption due to increased vehicle movement and service volume, the net energy cost—the difference between charging expenses and discharging revenues—was cut by 50%. This dramatic saving is attributed to the model’s ability to anticipate price fluctuations and strategically schedule charging during low-cost periods while selling stored energy back to the grid during high-price windows. In contrast, the DQN model, while attempting to incorporate future value, failed to fully capture the temporal patterns of electricity pricing and often converged to local optima, leading to less effective energy management.
The success of the model also highlights the importance of holistic system design in autonomous mobility. Rather than treating routing, charging, and passenger matching as separate optimization problems, the integrated framework recognizes their interdependence. For example, sending a vehicle to a distant high-demand zone may incur immediate travel costs but can yield higher long-term rewards if it leads to multiple successive rides. Similarly, delaying a pickup to first recharge at a nearby station may slightly extend passenger wait time but ensures the vehicle remains operational and avoids costly emergency relocations later.
Moreover, the model demonstrates how autonomous fleets can serve as flexible energy resources within smart city infrastructure. By aggregating the storage capacity of dozens or hundreds of vehicles, fleet operators can participate in demand response programs, provide frequency regulation services, and help balance renewable energy integration. This dual role—as both transportation providers and grid-supporting assets—enhances the economic viability of autonomous mobility services while contributing to broader sustainability goals.
Another key strength of the approach is its adaptability to different urban contexts and operational requirements. The neural network-based value function can be retrained with local demand patterns, road networks, and electricity tariffs, making it applicable to cities worldwide. Furthermore, the modular design allows for the incorporation of additional constraints or objectives, such as battery degradation models, charging station congestion, or environmental impact metrics, without requiring fundamental changes to the core algorithm.
Despite its achievements, the authors acknowledge several limitations and avenues for future work. The current model assumes idealized conditions, including instantaneous charging/discharging, no queuing at charging stations, and perfect knowledge of future demand distributions. In real-world deployment, factors such as battery aging, variable charging rates, and uncertain rider behavior would need to be addressed. Additionally, the transition from simulation to real-world testing presents challenges related to sensor accuracy, communication latency, and safety assurance.
Nevertheless, the study represents a significant step forward in the intelligent management of autonomous electric fleets. Its emphasis on long-term value over short-term gains aligns with the operational realities of 24/7 mobility services, where decisions made today directly influence profitability tomorrow. As cities worldwide seek to reduce congestion, emissions, and transportation costs, solutions like this offer a pathway toward more efficient, sustainable, and economically viable urban mobility systems.
The implications extend beyond ride-hailing services. The same principles could be applied to other shared autonomous vehicle applications, such as delivery fleets, microtransit shuttles, or last-mile logistics. In each case, the ability to anticipate future demand and optimize both movement and energy use can lead to substantial performance improvements. Moreover, as autonomous vehicle technology matures and regulatory frameworks evolve, the integration of V2G capabilities could become a standard feature, transforming electric vehicles into integral components of distributed energy networks.
From a policy perspective, the research underscores the need for coordinated planning between transportation and energy sectors. Municipalities investing in autonomous mobility should consider the synergies with smart grid development, ensuring that charging infrastructure is not only sufficient in quantity but also strategically located and capable of bidirectional power flow. Incentive structures, such as dynamic pricing or congestion charges, can further guide fleet behavior toward socially optimal outcomes, such as reduced downtown congestion or lower peak electricity loads.
For industry stakeholders, the study provides a compelling business case for adopting advanced AI-driven dispatch systems. While the initial investment in data infrastructure and machine learning expertise may be significant, the long-term returns in terms of increased revenue, reduced operating costs, and enhanced customer satisfaction can far outweigh these expenses. Companies that embrace such technologies early may gain a competitive edge in the rapidly evolving mobility-as-a-service market.
In conclusion, the work by Zeng Weiliang, Han Yu, and Fu Hui presents a comprehensive and practical solution to one of the most pressing challenges in autonomous transportation: how to balance immediate service demands with long-term operational efficiency. By embedding future revenue expectations into real-time dispatch decisions and leveraging the dual functionality of electric vehicles as both transporters and energy carriers, the model sets a new standard for intelligent fleet management. As autonomous electric taxis move from pilot projects to widespread deployment, approaches like this will be essential for unlocking their full economic and environmental potential.
Zeng Weiliang, Han Yu, Fu Hui. Industrial Engineering Journal. doi: 10.3969/j.issn.1007-7375.230095