Optimal Rebalancing with Waiting Time Constraints for a Fleet of Connected Autonomous Taxi Seong Ping Chuah, Shili Xiang, Huayu Wu Institute for Infocomm Research, A*STAR 1 Fusionopolis Way, Singapore 138632 Email: {chuahsp,sxiang,huwu}@i2r.a-star.edu.sg Abstract—A ﬂeet of cooperative autonomous taxi is an emerg- ing application of IoT in transportation industry. Unlike manned taxis that cruise on roads uncoordinated and often compete for passengers, autonomous vehicle can move cooperatively to transport passengers more efﬁciently. In this paper, we present a case study on an IoT application of new cooperative management technique for a ﬂeet of autonomous taxi. In transportation network, optimal rebalancing allows sustainable ﬂow of vehicle with a minimum number of vehicle to transport passengers ﬂows in uneven directions. However, long waiting time to board a taxi during peak hours degrades quality of service. To tackle this issue, we extend recent advances in autonomous mobility-on- demand solution to incorporate waiting time policy. Speciﬁcally, we introduce stability and control of passenger’s queues in the optimal rebalancing to conﬁne the queues (thus waiting time in queues) to a speciﬁed range. We validate our new technique via data-driven simulations of a ﬂeet of autonomous taxi by leveraging on Singapore’s taxi dataset. Data-driven simulations demonstrate promising results of the new technique in ensuring efﬁcient and low waiting time of taxi service for passengers. I. I NTRODUCTION While vehicle-to-vehicle communication systems enables vehicles to ”talk” to each other, they are still manned by human drivers who make every decisions on road. Emerging technology in autonomous vehicle has signiﬁcantly expanded the application of internet-of-thing (IoT) to revolutionizes the way people transit in town [1]. A ﬂeet of internet-connected autonomous taxi can communicate and make decisions coop- eratively to cater for large urban mobility needs in much more efﬁcient way [2]. Traditional manned taxis cruise in an uncoordinated manner, and often compete among each other for passengers. Many schemes, such as [3], [4], [5], [6], [7], [8] have been proposed to help taxi drivers match with passenger demands. More recently, Xu [9] investigated social propagation effects in predicting taxi drivers’ future behaviors. All these schemes, however, play only an advisory role to the taxi drivers while cooperation level among taxis remains low. Deploying autonomous vehicles for taxi service gives rise to a new challenge in cooperative management of the au- tonomous vehicles on roads[10] for efﬁcient transportation of passengers. In autonomous mobility-on-demand (AMOD) ser- vice, the autonomous vehicles coordinate with the command center on their heading directions while handling the local driving task autonomously. With optimal coordinating policy, AMOD can be more efﬁcient in serving the mobility needs of passengers with minimal number of vehicle [10]. On the other hand, passenger’s trafﬁcs often ﬂow in uneven directions during peak hours. [11] Rebalancing routes the vehicles from the destinations to the sources of trafﬁcs to allow sustainable ﬂow of vehicle to transport passengers. In [10], [12], AMOD service was modeled in queuing theoretical framework, where autonomous taxis are assumed to pick- up/drop-off passengers at a number of stations. Optimal re- balancing policy was then formulated to serve the passenger’s ﬂows with minimal rebalancing trafﬁcs. However, passengers’ waiting times and the build-up of queues [13], [14] at stations were not explicitly captured, analyzed and constrained. Cooperative management of autonomous taxi ﬂeet improves the service efﬁciency, and is crucial to the successful de- ployment of this new IoT application. However, the lack of operational trials and data from true autonomous taxi ﬂeet poses challenges in ﬂeet management and optimization. In this paper, we leverage on dataset from existing manned taxis in Singapore to conduct data-driven simulations, and validate the new cooperative management technique for AMOD deploy- ment. Simulations based on real-world dataset render us more conﬁdence for future deployment of AMOD service. The paper is organized as follows. We describe the queuing theoretical framework in Section II. In Section III, we for- mulate the optimal rebalancing with waiting time policy as an optimization. We present the solution procedure in IV. We showcase a Singapore’s case study in Section V. The paper is concluded in Section VI II. MODEL DESCRIPTION We consider a ﬂeet of autonomous taxi roaming on roads to provide transportation service to passengers. Let there be a set N of stations with substantial demand for taxi service. These stations can be identiﬁed as the active points of taxi pick-up/drop-off via clustering of taxi rides within the area. Passenger demands arrive at station i ∈N according to a Poisson arrival process with rate λ i , and request for taxi service destined to another station j ∈N with the probability of p ij where p ij ∈ R, ∑ j p ij =1,i = j,p ii =0 ∀ i ∈N . Upon arrival at a station, the passenger takes the taxi service if autonomous vehicles are available at the station. If no autonomous vehicle is available in the station, the passenger queue to wait for taxi ride in ﬁrst-in-ﬁrst-out manner. Upon