Online Transportation Problem Overview

• Online Transportation Problem is an online assignment challenge where requests are irrevocably matched to static servers with capacity limits in a metric space.
• The topic generalizes online metric matching by incorporating multiple server capacities and optimizing assignment costs, leading to competitive algorithm designs.
• Recent research explores diverse variants including multi-robot routing, time-dependent vehicle scheduling, and hybrid online-offline booking systems.

Searching arXiv for papers on the Online Transportation Problem and closely related formulations. The Online Transportation Problem (OTP) is an online assignment problem in which requests arrive over time and must be assigned immediately and irrevocably to available service resources under metric costs and capacity constraints. In the classical metric formulation, a set of servers is clustered at server sites in a metric space, each site has an integer capacity, and each arriving request must be matched to one free site so as to minimize total transportation cost. The problem generalizes online metric matching by allowing multiple capacities per site, yet it differs fundamentally from moving-server models such as the $k$-server problem because OTP assigns static capacity without movement costs and seeks to minimize transport distances from requests to server sites (Harada et al., 2024). More recent work also uses the term in broader operational settings, including online pickup-and-delivery routing, communication-aware multi-robot transportation, and hybrid online-offline booking systems; these variants retain the core online feature that decisions are made under partial or sequentially revealed information, but they optimize richer objectives such as makespan, tardiness, or route cost rather than only assignment distance (Liao et al., 25 Mar 2026, Kontogiannis et al., 2024, Sivagnanam et al., 2022).

1. Classical metric formulation and objective

In the classical formulation studied in recent theoretical work, one is given a metric space $(X,d)$ satisfying the triangle inequality. The online transportation problem $\mathrm{OTR}(k,m)$ has $k$ servers clustered at $m$ server sites $S \subseteq X$, where each site $s \in S$ has capacity $c(s) \in \mathbb{N}$ with $c(s) \ge 1$ and $\sum_{s \in S} c(s)=k$ (Harada et al., 2024). Requests are unit demands arriving online, at most $(X,d)$0 in total, and each request must be assigned immediately and irrevocably to one free server site; splitting a single request across multiple servers is not allowed. If request $(X,d)$1 is assigned to site $(X,d)$2, the incurred cost is $(X,d)$3, and total cost is the sum over all requests. The objective is to minimize this total assignment cost (Harada et al., 2024).

A special case is Online Metric Matching $(X,d)$4, obtained when $(X,d)$5 and every site has capacity $(X,d)$6; a further restriction assumes that every request arrives exactly at a server location, i.e. $(X,d)$7 (Harada et al., 2024). The offline optimum is the optimal integral solution of a transportation linear program with primal variables $(X,d)$8, objective

$(X,d)$9

request-assignment constraints $\mathrm{OTR}(k,m)$0, and site-capacity constraints $\mathrm{OTR}(k,m)$1. The corresponding dual uses request variables $\mathrm{OTR}(k,m)$2 and nonnegative capacity variables $\mathrm{OTR}(k,m)$3, with constraints $\mathrm{OTR}(k,m)$4 (Harada et al., 2024). Competitive analysis compares an online algorithm $\mathrm{OTR}(k,m)$5 against this offline optimum via

$\mathrm{OTR}(k,m)$6

The central deterministic lower bound for $\mathrm{OTR}(k,m)$7 server sites is

$\mathrm{OTR}(k,m)$8

which follows from the online metric matching lower bound by clustering servers at $\mathrm{OTR}(k,m)$9 sites (Harada et al., 2024). This lower bound has framed the modern theory of OTP, especially the long-standing conjecture of Kalyanasundaram and Pruhs that a deterministic $k$0-competitive algorithm exists (Harada et al., 2024).

2. Structural relation to online matching, routing, and transportation systems

OTP sits at an intersection of online matching, capacitated assignment, and transportation routing, but its modeling assumptions vary significantly across subliteratures. In the strict metric-assignment literature, the resources are static server sites with capacities and the only cost is request-to-site distance. Capacities create fundamentally different challenges from unit-capacity matching, because naive policies can prefer filling certain sites first and suffer exponential losses in $k$1 (Harada et al., 2024). This is why OTP is not merely OMM with aggregated servers.

A distinct but related use of the term appears in online vehicle-routing and logistics formulations. In time-dependent pickup-and-delivery routing, requests are tuples with pickup and delivery nodes, service times, deadlines, loads, and vehicle-type eligibility, and online decisions insert each arriving request into an evolving route while preserving capacity and temporal feasibility (Kontogiannis et al., 2024). In a smart-factory transportation multi-robot system, the factory floor is modeled as an undirected graph $k$2, AGVs act under partial observability, and tasks arise dynamically with precedence indices, due times, payload constraints, congestion thresholds, and communication-dependent global state estimation (Liao et al., 25 Mar 2026). In paratransit, a hybrid formulation called Offline VRP with Online Bookings requires confirming tight pickup windows in real time while the full routing problem is solved offline after bookings are known (Sivagnanam et al., 2022).

These formulations share the online, irrevocable, and resource-constrained character of OTP, but they optimize different quantities. Classical metric OTP minimizes assignment distance (Harada et al., 2024). The smart-factory formulation minimizes makespan

$k$3

subject to task assignment, tardiness, payload, and occupancy constraints (Liao et al., 25 Mar 2026). The time-dependent routing formulation prioritizes serving as many requests as possible and secondarily minimizing global service cost such as total travel time or total distance (Kontogiannis et al., 2024). The paratransit formulation optimizes the downstream offline VRP cost induced by online commitments to tight pickup windows (Sivagnanam et al., 2022). A plausible implication is that “Online Transportation Problem” functions as a family resemblance term across communities rather than a single universally fixed formalism.

3. Historical bounds, prior algorithms, and resource augmentation

The historical baseline for deterministic online metric matching is sharp: for $k$4, the deterministic lower bound is $k$5, with a matching upper bound by Permutation on OMM due to Kalyanasundaram–Pruhs and Khuller–Mitra–Vishkin (Harada et al., 2024). For capacitated OTP, however, upper bounds long depended poorly on the number of sites. Reported prior results include Greedy with competitive ratio approximately $k$6 and Permutation with $k$7 competitive ratio when capacities are present (Harada et al., 2024). Nayyar–Raghvendra later showed Robust-Matching to be $k$8-competitive, tightening the upper bound to $k$9 (Harada et al., 2024).

A parallel line of analysis studies the natural greedy rule under resource augmentation. In that setting, parking garages or server sites have original capacities $m$0, while the online algorithm is given augmented capacities $m$1 for $m$2, and the offline optimum is still evaluated with the original capacities (Arndt et al., 2023). The greedy rule assigns each arriving request to the nearest garage with remaining augmented capacity. For $m$3, the paper proves

$m$4

with examples such as $m$5, $m$6, and $m$7 as $m$8 (Arndt et al., 2023). The bound is essentially tight: for every $m$9, there exists an instance with ratio exceeding $S \subseteq X$0 (Arndt et al., 2023). This line does not resolve the unaugmented deterministic OTP, but it shows that plain greedy becomes constant-competitive once capacities are multiplied by at least three.

The progression of results clarifies the conceptual divide between two questions. One asks whether unaugmented deterministic OTP admits the conjectured $S \subseteq X$1-competitive algorithm. The other asks how much extra capacity suffices to make simple online rules near-optimal. The 2023 resource-augmentation analysis answers the second affirmatively for greedy (Arndt et al., 2023), while the 2024 deterministic algorithm answers the first partially by giving the first $S \subseteq X$2-competitive deterministic algorithm without extra resources (Harada et al., 2024).

4. Subtree-Decomposition and the first $S \subseteq X$3-competitive deterministic algorithm

The paper "A Nearly Optimal Deterministic Algorithm for Online Transportation Problem" introduces Subtree-Decomposition, a new deterministic algorithm for OTP and proves that it achieves

$S \subseteq X$4

for $S \subseteq X$5, which is the first $S \subseteq X$6-competitive deterministic bound and comes within a constant factor of the lower bound $S \subseteq X$7 (Harada et al., 2024). The algorithm is developed first on power-of-two weighted tree metrics, where every edge weight is $S \subseteq X$8 for some integer $S \subseteq X$9, and then extended to general metrics via a rounded minimum spanning tree embedding (Harada et al., 2024).

The core design uses a recursive decomposition of a rooted tree into subtrees at two scales. First, heavy versus light structure is captured by a maximal subtree $s \in S$0 containing the root and avoiding maximum-weight edges, together with child subtrees $s \in S$1 attached through those maximum-weight edges. Second, the whole tree is split into two coarse subtrees $s \in S$2 and $s \in S$3 by removing one root edge (Harada et al., 2024). The algorithm is inductively defined and belongs to the MPFS class, short for Most Preferred Free Servers: for each request and current set of free sites, it defines a deterministic priority order and picks the highest-priority free site. Its preference list $s \in S$4 depends only on the request and the tree decomposition, not on the arrival sequence or capacities (Harada et al., 2024).

The online rule operates in two phases. While $s \in S$5 still contains a free server, requests are first served within their native subtree if possible; otherwise the request is routed into $s \in S$6 by a proxy request at the attachment point $s \in S$7. After $s \in S$8 becomes full, the rule first tries the same fine subtree $s \in S$9, then the same coarse subtree $c(s) \in \mathbb{N}$0, and finally the other coarse subtree $c(s) \in \mathbb{N}$1 via proxy at the opposite coarse root (Harada et al., 2024). This phase transition is the mechanism that controls when and how load crosses heavy cuts.

Implementation is relatively direct on trees. The tree is rooted, $c(s) \in \mathbb{N}$2, $c(s) \in \mathbb{N}$3, and the $c(s) \in \mathbb{N}$4 are computed by a single DFS, and the coarse split $c(s) \in \mathbb{N}$5, $c(s) \in \mathbb{N}$6 is determined by removing one child-edge of the root. Each arrival requires only a constant number of proxy calls and a bounded number of subtree-availability checks. The assignment can be implemented in time proportional to the depth of the relevant subtree chain; in worst-case tree depth this is $c(s) \in \mathbb{N}$7, and with natural hierarchical data structures it is $c(s) \in \mathbb{N}$8 (Harada et al., 2024).

The competitiveness proof uses a hybrid algorithm framework for MPFS algorithms, specialized from Gupta–Lewi’s lemma. For a hybrid instance $c(s) \in \mathbb{N}$9, one bounds the hybrid cycle length $c(s) \ge 1$0; if $c(s) \ge 1$1 for every hybrid, then the algorithm is $c(s) \ge 1$2-competitive (Harada et al., 2024). On power-of-two trees, the key bound is

$c(s) \ge 1$3

where $c(s) \ge 1$4 (Harada et al., 2024). Combined with a monotonicity property of Subtree-Decomposition, this yields

$c(s) \ge 1$5

for $c(s) \ge 1$6, the online matching problem on the power-of-two tree (Harada et al., 2024). General metrics are then handled by computing a minimum spanning tree $c(s) \ge 1$7 on server locations, rounding edge weights up to powers of two, and using the distortion facts $c(s) \ge 1$8 and $c(s) \ge 1$9. This produces the sequence of bounds

$\sum_{s \in S} c(s)=k$0

for OMM with requests at server locations,

$\sum_{s \in S} c(s)=k$1

for general OMM via the Meyerson–Naor–Paz server-equivalence reduction, and finally

$\sum_{s \in S} c(s)=k$2

for OTR via the MPFS reduction from $\sum_{s \in S} c(s)=k$3 to $\sum_{s \in S} c(s)=k$4 (Harada et al., 2024).

The constants arise from identifiable steps: $\sum_{s \in S} c(s)=k$5 on power-of-two trees, a factor below $\sum_{s \in S} c(s)=k$6 from MST rounding, a factor $\sum_{s \in S} c(s)=k$7 from server-equivalence, and then replacement of $\sum_{s \in S} c(s)=k$8 by $\sum_{s \in S} c(s)=k$9 for capacitated sites (Harada et al., 2024). This decomposition also isolates the main remaining obstacles to the conjectured $(X,d)$00: avoiding the distortion from rounded MST embeddings and improving the server-equivalence reduction.

5. Variants in routing, robotics, and multimodal logistics

Outside the strict metric-assignment model, OTP-like problems often involve routing, temporal coupling, or communication constraints. In VRPPDSTCtd, online requests are pickup-and-delivery jobs on a directed road network $(X,d)$01 with time-dependent, vehicle-type-specific arc travel-time functions $(X,d)$02, continuous piecewise linear structure, and the FIFO property, which implies nondecreasing arc-arrival-time functions $(X,d)$03 (Kontogiannis et al., 2024). The paper proposes two online schedulers: TDINSERTION, a time-dependent variant of Plain-Insertion, and TDPROPHET, which augments insertion with short-term demand forecasts. Feasibility checking requires updating time-dependent arrival and departure labels on route suffixes, yielding $(X,d)$04 time to enumerate feasible insertion pairs along a worker route (Kontogiannis et al., 2024). The work does not provide competitive guarantees, but it formalizes online transportation as online route insertion with pickup-before-delivery precedence, capacities, and exact time windows.

In smart factories, OTP is formulated as online multi-robot task assignment plus congestion-aware route scheduling under partial observability (Liao et al., 25 Mar 2026). The factory is an undirected graph $(X,d)$05; AGVs have local sensing radius $(X,d)$06, payload states, and action space $(X,d)$07. The true state $(X,d)$08 includes positions, intentions, tasks, and production readiness, while AGV $(X,d)$09 observes

$(X,d)$10

Task assignment uses a binary matrix $(X,d)$11, and route planning is performed in a spatiotemporal search space with congestion-aware A* cost

$(X,d)$12

where congestion penalties are derived from a global map $(X,d)$13 (Liao et al., 25 Mar 2026). The overall objective is makespan minimization under tardiness, payload, and occupancy constraints, including $(X,d)$14 for congestion-free occupation (Liao et al., 25 Mar 2026).

The same paper couples scheduling with wireless M2M communication. AGVs transmit current state $(X,d)$15 and navigation route $(X,d)$16, the edge server constructs and broadcasts a global congestion map $(X,d)$17, and uplink reliability is improved via retransmission-free multi-link transmission over $(X,d)$18 orthogonal channels (Liao et al., 25 Mar 2026). The design is explicitly scheduling-oriented: the communication interval $(X,d)$19 and number of selected channels $(X,d)$20 that minimize makespan differ from those that maximize throughput $(X,d)$21, which the paper presents as evidence that M2M communication for AGV scheduling fundamentally differs from H2H system design (Liao et al., 25 Mar 2026). This suggests a broadening of OTP from pure online assignment to communication-computation-control co-design.

A different extension appears in multimodal container assignment, where each arriving container must be assigned online to a train or a truck subject to departure-time, arrival-time, and train-capacity constraints (Farahani et al., 2021). The paper formulates the objective

$(X,d)$22

and uses a Deep Q-Network with input size $(X,d)$23, two hidden layers of 100 nodes each, ReLU activations, Adam optimization, and action masking

$(X,d)$24

to enforce feasibility (Farahani et al., 2021). Here, OTP becomes an online multimodal assignment problem with learned sequential decision-making rather than worst-case competitive analysis.

6. Hybrid and continuous formulations beyond one-shot assignment

Some OTP-adjacent models explicitly bridge online commitment and offline optimization. In Offline Vehicle Routing Problem with Online Bookings, requests are not routed immediately when they arrive; instead, each booking arrives with a broad pickup window $(X,d)$25, and the agency must confirm in real time a tight window $(X,d)$26 with $(X,d)$27 (Sivagnanam et al., 2022). After all bookings are received, routes are optimized offline as a pickup-and-delivery VRP with time windows, capacity, and ride-time constraints. The coupling is formalized through a policy $(X,d)$28 that chooses $(X,d)$29 online while an anytime solver $(X,d)$30 continuously improves a warm-start route plan between calls (Sivagnanam et al., 2022). This hybrid captures a setting in which the online decision is not dispatch but commitment.

The computational approach combines a learning-based policy with an anytime heuristic. The booking-time policy is a simplified deep Q-network that evaluates candidate tight windows using a shaped immediate cost

$(X,d)$31

while the anytime solver alternates greedy construction and simulated annealing between calls (Sivagnanam et al., 2022). A plausible implication is that OTP methodologies can be organized by the locus of irreversibility: immediate assignment, immediate route insertion, or immediate service-window commitment.

At the other end of the spectrum is continuous transportation in multi-agent systems. In that model, a set of agents repeatedly visits fixed locations to pick up objects replenished by independent Poisson processes and deliver them to a hub $(X,d)$32, maximizing long-run throughput

$(X,d)$33

rather than minimizing one-shot assignment cost (Wang et al., 2015). Distances are metric, travel time for agent $(X,d)$34 from $(X,d)$35 to $(X,d)$36 is

$(X,d)$37

and replenishment rates $(X,d)$38 are initially unknown and learned online (Wang et al., 2015). The proposed Online Balanced Partitioning algorithm starts from $(X,d)$39-means clusters, estimates replenishment rates online, and triggers repartitioning when cluster loads become imbalanced. Within each cluster, an agent uses a Greedy Rate rule based on expected pickup per unit travel time (Wang et al., 2015). Although this model is structurally different from capacitated metric OTP, it preserves online transportation’s essential features of uncertainty, capacity limits, and irrevocable motion decisions.

7. Empirical performance, misconceptions, and open directions

A common misconception is that OTP is adequately captured by nearest-neighbor or greedy assignment. In the classical capacitated setting, this is false without augmentation: capacities fundamentally alter the problem, and greedy-like policies can incur exponential loss in the number of sites (Harada et al., 2024). The 2023 augmentation result shows only that greedy becomes constant-competitive when the online algorithm has $(X,d)$40 times the site capacities, not in the original-capacity model (Arndt et al., 2023). Another misconception is that OTP is simply the $(X,d)$41-server problem in static disguise. In fact, the $(X,d)$42-server problem incurs movement costs and maintains server states, whereas OTP assigns static capacity without movement costs (Harada et al., 2024).

Empirically, OTP-style methods have produced strong results in application-driven variants. In the smart-factory AGV setting, communication-enabled scheduling achieved a 54.42% improvement over local reasoning at $(X,d)$43, and under the severe resource-constraint setting $(X,d)$44, $(X,d)$45, it yielded a 17.27% efficiency gain versus local reasoning, with experiments repeated five times and averaged (Liao et al., 25 Mar 2026). In time-dependent online vehicle routing, TDINSERTION variants with distance metric and $(X,d)$46 objective showed 14.6\% to 49.1\% decrease in total-length and 13.5\% to 48.9\% decrease in total travel-time relative to human-curated assignments, while TDPROPHET added roughly 4\% further improvement (Kontogiannis et al., 2024). In multimodal container assignment, the DRL method outperformed tested competitors by 20.48\% to 55.32\% in total transportation cost and by 7.51\% to 20.54\% in train-capacity utilization, while staying within 2.7\% of offline ILP optimal cost and 0.72\% of optimal capacity utilization (Farahani et al., 2021). In continuous transportation, Online Balanced Partitioning outperformed GR+EO and random transportation with statistically significant differences, reported as $(X,d)$47 versus GR+EO (Wang et al., 2015). These findings belong to distinct formulations, but together they indicate that online transportation problems are increasingly treated as integrated systems problems rather than isolated matching instances.

For the classical deterministic OTP, the principal open problem remains the Kalyanasundaram–Pruhs conjecture of a deterministic $(X,d)$48-competitive algorithm (Harada et al., 2024). The 2024 result narrows the gap to a constant factor, but the analysis identifies two major sources of loss: the factor below $(X,d)$49 introduced by rounding MST edge weights to powers of two, and the $(X,d)$50 server-equivalence reduction from requests at server locations to general OMM (Harada et al., 2024). The paper explicitly asks whether these constants can be improved, whether a direct analysis for general metrics can avoid tree embedding, and whether randomized algorithms for general metrics and capacities can achieve $(X,d)$51 or even $(X,d)$52 (Harada et al., 2024). In application-driven variants, future directions include distributed multi-agent reinforcement learning for adaptive collision resolution and deeper integration of communication, computation, and control in smart factories (Liao et al., 25 Mar 2026), as well as richer forecasting and metaheuristics in time-dependent routing (Kontogiannis et al., 2024).

Taken together, these developments show that the Online Transportation Problem has become both a sharply defined object in online algorithms and a broader organizing paradigm for online logistics, routing, and robotic transportation. In its narrowest sense, it is the capacitated online metric assignment problem with immediate irrevocable decisions and distance-minimization objective, now known to admit a deterministic $(X,d)$53-competitive algorithm (Harada et al., 2024). In its broader sense, it denotes a class of online transportation systems in which uncertainty, capacities, temporal coupling, and partial information force decisions to be made before the full future is known.