Semi-on-Demand Transit Feeder Service

• Semi-on-demand transit feeder service is a design space that merges fixed-route schedules with on-demand segments to efficiently link low-density areas with trunk networks.
• It combines hybrid routes, on-demand shuttles, semi-flexible connectors, and ride-pooling to allocate flexibility where fixed service is inefficient.
• Simulation, optimization, and reinforcement learning are key tools used to balance accessibility, generalized cost, and fleet efficiency in these systems.

Semi-on-demand transit feeder service denotes a family of feeder operations that connect travelers to a trunk public transport network by combining a structured service component with a flexible one. In the recent literature, this family includes hybrid routes that serve scheduled fixed stops before entering a flexible area, on-demand shuttles synchronized with high-frequency bus or rail trunks, semi-flexible connectors that follow predefined swaths while responding to requests en route, and ride-pooling or meeting-point feeders organized around stations or hubs. Across these variants, the central design problem is the same: allocate flexibility where low density makes fixed service inefficient, while preserving the line-haul productivity, transfer structure, and predictability of the main network (Ng et al., 2024, Ng et al., 2024, Auad et al., 2021, Zhen et al., 2024).

1. Conceptual scope and service archetypes

Semi-on-demand feeders are positioned between conventional fixed-route service and fully on-demand microtransit. In the simulation literature, fixed-route, semi-flexible, and on-demand operations are treated as distinct design classes, with the semi-flexible class retaining a fixed skeleton while allowing deviations or virtual-stop service under explicit slack, backtracking, and detour constraints (Yoon et al., 2021). In multimodal formulations, the feeder function is first-mile and last-mile access to high-frequency trunks, typically rail or bus rapid transit, with transfer limits and synchronization treated as design variables rather than operational afterthoughts (Auad et al., 2021).

At the architectural level, several recurring patterns appear. Hybrid-route services use a fixed segment in denser areas and a flexible segment in lower-density areas; this is the dominant formulation in the Shared Autonomous Mobility Service and Shared Autonomous Vehicle literature (Ng et al., 2024, Ng et al., 2024). On-Demand Multimodal Transit Systems integrate on-demand shuttles with high-frequency buses and trains so that passengers are picked up near their origins, delivered to a hub, moved on the trunk network, and in some cases distributed by a final shuttle at the destination end (Auad et al., 2021). Semi-flexible connectors differ from fully-flexible connectors in that the vehicle follows a predefined path through a swath and serves requests received en route, rather than solving a closed set of pre-dispatch requests through an optimal tour (Zhen et al., 2024). A distinct feeder-network design line models multi-circulation feeder loops that start and end at the same rail or BRT station and are generated as contiguous circular or separated routes over an urban road network (Sani et al., 2022).

Form	Operating logic	Representative study
Hybrid route	Fixed segment in high-density area, flexible segment in low-density area	(Ng et al., 2024)
ODMTS feeder	On-demand shuttle linked to high-frequency bus/rail trunk	(Auad et al., 2021)
Semi-flexible connector	Predefined swath/path with requests served en route	(Zhen et al., 2024)
Circular feeder loop	Continuous loop starting and ending at the same station	(Sani et al., 2022)
Ride-pooling feeder	Shared feeder trips pooled before or around dispatch	(Fan et al., 2024)

This variety implies that “semi-on-demand feeder service” is best understood as a design space rather than a single canonical model. A plausible implication is that comparisons across studies are most meaningful when they control for corridor geometry, demand density, transfer architecture, and the degree to which flexibility is localized to zones, swaths, meeting points, or full door-to-door pickup.

2. Demand representation, generalized cost, and behavioral modeling

The analytical core of semi-on-demand feeder design is almost always a generalized-cost formulation that combines user and operator terms. Hybrid-route directional-service models explicitly separate access, waiting, and riding costs for users from distance-based operating and time-based vehicle costs for operators, and optimize the length of the flexible portion, fleet size, headway, and vehicle size (Ng et al., 2024). Off-peak SoD planning for SAVs uses the same decomposition, with user cost

$$c^u_r = \gamma^a t^a_r + \gamma^w t^w_r + \gamma^r t^t_r$$

and generalized system cost

$$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$

so that access reduction, waiting-time changes, detours, and operator mileage can be assessed in a single objective (Ng et al., 2024). Reinforcement-learning-based zonal SoD control adopts the same cost structure, written as

$$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$

thereby embedding dispatching control in a system-cost framework rather than a purely operational one (Ng et al., 2 Sep 2025).

Demand is represented in several complementary ways. Continuous-approximation models for adaptive multimodal transit use an exponentially decaying radial density,

$p(x, t) = p_0(t) \cdot e^{-\gamma x},$

to decide where fixed-route feeders or demand-responsive feeders should operate at different times of day (Calabro' et al., 2021). In SimFLEX, the feeder-versus-public-transport decision is modeled by a binary logit,

$$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$

with utilities built from travel, wait, and walk times, transfers, fares, operational parameters, and behavioral coefficients; expected feeder travel times are then updated through a Method of Successive Averages learning process until convergence (Vasiutina et al., 24 Apr 2025). Demand-agnostic pre-deployment assessment replaces observed demand with a fraction of demand, denoted by $\alpha$, interpreted as the proportion of residents potentially interested in the service, and varies $\alpha$ over a range such as $0.1\%$ to $5\%$ to identify which area-hub pairs cross KPI thresholds first (Shulika et al., 16 Jun 2026).

A distinct feeder-design tradition focuses on potential demand over links and stations. The modified potential demand model for multi-circulation feeder networks defines link attractiveness using traffic demand on each link, average walking distance, distance to candidate stations, demand for accessing each main station from the link, and a three-level discrete connection-interest index $g_{ijs} \in \{1,2,3\}$ for low, medium, and high connection strength (Sani et al., 2022). This makes station preference endogenous to route design rather than an external assignment step.

3. Network design, routing, and control algorithms

Semi-on-demand feeder systems are planned and operated through a broad algorithmic spectrum ranging from closed-form continuous approximation to mixed-integer programming and learning-based control. In network design for circular feeders, the objective is to maximize captured potential demand,

$$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$0

subject to route-allocation, travel-time, and circularity constraints. A labelling constraint of the form

$$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$1

is used to ensure contiguous circuits, and the solution procedure combines a constructive heuristic inspired by Clarke-Wright with a genetic algorithm whose genes are routes and whose chromosomes are entire feeder networks (Sani et al., 2022). This is a route-generation paradigm for structured feeder loops rather than dynamic microtransit.

In multimodal on-demand feeder design, the core network-design problem is often written on a transfer-expanded graph. ODMTS formulates network design on a directed multigraph with bus-arc activation variables $$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$2 and trip-flow variables $$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$3, minimizing fixed operational cost plus passenger-routing cost under a bound on the number of legs. Transfer limits are enforced by graph expansion instead of explicit combinatorial constraints, and Benders decomposition is used to separate fixed-route design from passenger routing. Fleet sizing is then handled by ridesharing optimization, a set partitioning model, and a min-flow covering problem, while real-time dispatching is implemented through a rolling-horizon Dial-A-Ride procedure (Auad et al., 2021).

Connector design reveals a second major algorithmic divide: fully-flexible versus semi-flexible routing. For fully-flexible routing, local pickup tours are represented by

$$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$4

whereas semi-flexible swath service uses

$$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$5

The refined analytical models for these two strategies show that the transition from fully-flexible to semi-flexible routing occurs as demand and region size increase, and that zoning geometry differs structurally: fully-flexible routing favors square zones, while semi-flexible routing prefers elongated rectangular zones (Zhen et al., 2024).

At the operational level, ride-pooling feeders and zonal SoD services introduce explicitly adaptive control. Ride-Pooling as Feeder develops a batch-based matching algorithm, an adaptive dispatching algorithm with occupancy targets, an urgency-based repositioning rule, and open-TSP routing for small pickup batches. The analytical tuning result

$$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$6

links the optimal batch size and matching buffer directly to the occupancy target and outbound request density (Fan et al., 2024). RL-based zonal dispatching instead formulates feeder control as a Markov Decision Process and uses Proximal Policy Optimization. The state includes fleet availability, demand forecast, and zone-level operational variables; the action space assigns SAVs to the regular route, to a flexible zone, or to holding; and the reward is the negative of unserved requests (Ng et al., 2 Sep 2025).

4. Simulation, benchmarking, and ex ante assessment

Because semi-on-demand feeder performance depends strongly on stochastic demand, geometry, and control rules, simulation is central to the field. The open-source simulation sandbox for fixed-route, semi-flexible, and on-demand designs uses a discrete-time agent-based framework with controllable service configuration, operational constraints, and contextual parameters, and evaluates ridership, weighted travel time, vehicle miles traveled, and rejection rate (Yoon et al., 2021). In that framework, the semi-flexible Mobility Allowance Shuttle Transit system uses a fixed checkpoint skeleton plus deviation slack, maximum backtracking, and walking-to-pickup options, which makes it particularly suitable as a benchmark feeder architecture rather than a one-off routing heuristic.

SimFLEX extends ex ante feeder evaluation toward site selection. It combines spatial, demographic, and GTFS-based transport data; uses OpenTripPlanner for public-transport routing and ExMAS for shared-trip attributes; generates multiple demand replications; and models day-to-day traveler learning through the Method of Successive Averages. After stabilization, it computes KPIs such as attractiveness, waiting-time reduction, added value, vehicle occupancy, and travel-hour indicators, and then repeats the analysis across a range of alternative-specific constants to test robustness of feeder-area ranking (Vasiutina et al., 24 Apr 2025). The demand-agnostic Krakow framework pushes this further by assuming demand is not yet known, generating requests from residential address points for alternative values of $$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$7, matching them by ExMAS, and selecting the optimal area-hub combination as the one that first satisfies thresholds such as $$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$8, $$c^g = \sum_{r \in \mathcal{R}} c^u_r + \sum_{v \in \mathcal{V}} c^o_v$$9, and $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$0 (Shulika et al., 16 Jun 2026).

Large-scale multimodal simulation and Pareto analysis provide a system-level benchmark for feeder roles. The bi-modal Berlin-Brandenburg study couples rail-bound line service with on-demand shuttles in MATSim and evaluates service quality, relative energy use, pooling efficiency, traffic volume, waiting time, detour, occupancy, and train loading under user-adoption levels of $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$1 and $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$2 (Sharma et al., 2023). The result is not merely a comparison of feeder policies but an explicit energy-quality frontier for multimodal feeder integration.

5. Empirical findings and operating regimes

Case studies show that semi-on-demand feeder performance is highly context dependent but consistently shaped by the same accessibility–detour–fleet trade-off. In Tehran District 10, the modified potential-demand model produced four feeder routes with lengths of $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$3, $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$4, $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$5, and $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$6 km, each starting and ending at the same railway station; the resulting network covered $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$7 of the area with a maximum access distance of $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$8 m, and accessibility rose from $$c^G = \sum_{r \in \mathcal{R}} (\gamma^A t^A_r + \gamma^W t^W_r + \gamma^T t^T_r) + \sum_{v \in \mathcal{V}} (\gamma^O_v d_v + \gamma^V_v t^V_v),$$9 to $p(x, t) = p_0(t) \cdot e^{-\gamma x},$0 (Sani et al., 2022). In Chicago, the analytical hybrid-route SAMS design found that $p(x, t) = p_0(t) \cdot e^{-\gamma x},$1 of feeder routes become hybrid, covering $p(x, t) = p_0(t) \cdot e^{-\gamma x},$2 of zones but only $p(x, t) = p_0(t) \cdot e^{-\gamma x},$3 of demand, and that hybrid feeder routes can reduce first/last-mile generalized cost by up to $p(x, t) = p_0(t) \cdot e^{-\gamma x},$4 for users in flexible areas and by about $p(x, t) = p_0(t) \cdot e^{-\gamma x},$5 network-wide, while lowering total generalized costs by $p(x, t) = p_0(t) \cdot e^{-\gamma x},$6 on average even though operator costs rise by about $p(x, t) = p_0(t) \cdot e^{-\gamma x},$7 (Ng et al., 2024).

Atlanta’s ODMTS study emphasizes resilience and multimodal coordination. For the pre-pandemic case, ODMTS delivered trips about $p(x, t) = p_0(t) \cdot e^{-\gamma x},$8 minutes shorter than the traditional system for most users excluding pure rail riders, with $p(x, t) = p_0(t) \cdot e^{-\gamma x},$9 shuttles and $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$0 buses. Average waits on most shuttle, bus, and rail legs were under $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$1–$$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$2 minutes, and more than $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$3 of shuttle and rail rides waited less than $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$4 minutes. Under pandemic scenarios with demand at $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$5 and $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$6 of normal, and even under a strict late-pandemic case with zero buses, rail at $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$7 capacity, and shuttles at one passenger, the system maintained service quality by scaling shuttle numbers to a range of $$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$8–$$P_{F_i} = \frac{\exp{ \mu U_{F_i} } }{ \exp{ \mu U_{F_i} } + \exp{ \mu U_{PT_i} } },$$9 (Auad et al., 2021).

Evaluation studies in Krakow and Munich make the trade-off structure especially explicit. SimFLEX found that Skotniki had higher service attractiveness by up to $\alpha$0 and added value by up to $\alpha$1, whereas Bronowice had a nearly $\alpha$2 greater potential for waiting-time reduction; the robustness analysis across alternative-specific constants still identified Skotniki as the stronger implementation candidate (Vasiutina et al., 24 Apr 2025). In Munich, off-peak SoD planning with SAVs reported generalized-cost reductions of $\alpha$3–$\alpha$4 relative to regular fixed route, with door-to-door convenience attracting more passengers when flexible-route lengths were moderate. The route-level example reported a reduction from about €$\alpha$5 to about €$\alpha$6 per passenger and an increase in served requests from $\alpha$7 to $\alpha$8 on one route (Ng et al., 2024). RL-based zonal control on another Munich route served $\alpha$9 more passengers at $\alpha$0 higher generalized costs on average than traditional fixed-route service, while the RL controller itself added $\alpha$1 more passengers at $\alpha$2 higher costs relative to nominal zonal control (Ng et al., 2 Sep 2025).

A recurring empirical result is that the preferred operating regime shifts with demand density. The autonomous minibus semi-on-demand study reported average generalized-cost reductions of \$\alpha$32.5 per passenger in grid networks, with selection indicators of $\alpha$4 for CTA Bus #126 and $\alpha$5 for CTA Bus #84, and effectiveness on the latter up to about $\alpha$6–$\alpha$7 passengers per hour (Ng et al., 2024). By contrast, in rural automated demand-responsive feeder transit with fixed fleet size, raising demand from $\alpha$8 to $\alpha$9 increased the circuitousness ratio from $0.1\%$0 to $0.1\%$1 and total cost per passenger from \$0.1\%$22.04%%%%7$p(x, t) = p_0(t) \cdot e^{-\gamma x},$7$0.1\%$5 and maximum average used capacity rose from $0.1\%$6 to $0.1\%$7 (Lee et al., 2019). Semi-on-demand therefore benefits from higher density only up to the point where detours, capacity pressure, or schedule slack begin to dominate.

6. Electrification, automation, and open design tensions

Recent work extends semi-on-demand feeders into electrified, autonomous, and modular regimes. Meeting-point-based electrified feeder services jointly optimize customer-to-meeting-point assignment, EV routing, and charging synchronization under limited charger capacity and partial recharging. Relative to a traditional door-to-door dial-a-ride benchmark, the optimized meeting-point system reduced fleet size by up to $0.1\%$8, in-vehicle travel time by $0.1\%$9, and vehicle kilometers by $5\%$0, while joint parameter configuration improved total kilometers traveled by $5\%$1 versus the benchmark (Ma et al., 2024). Mixed-fleet electrification planning generalizes this to transition periods by jointly choosing EVs, GVs, charger locations, charger types, and fleet size under stochastic demand and targeted CO$5\%$2 reduction. The case study in Bettembourg found that 24-seat EV operations were relatively insensitive to the number of chargers, whereas 10-seat EV systems required a minimum charger deployment before fleet downsizing became possible; rapid chargers provided the best balance relative to super-fast chargers, whose installation costs were harder to justify (Nakao et al., 17 Mar 2025).

Automation systematically shifts the feasible operating envelope toward smaller vehicles, shorter headways, and more flexible feeder regions. Adaptive Transit shows that lower crew costs make demand-responsive feeders optimal at shorter distances from the center and reduce walking cost by $5\%$3 compared with FRF-only design in automated-vehicle scenarios (Calabro' et al., 2021). The Munich SoD studies report that full SAV scenarios enable smaller vehicles, higher frequency, and larger flexible-route portions, whereas transition scenarios with a higher driver share impose higher headways and constrain the benefits of flexible-route operation (Ng et al., 2024). A more radical adjacent concept is Fly-by Transit, which uses autonomous mini-EV trailers for on-demand feeder access and platooned leaders for trunk movement; in a stylized corridor model it achieved more than $5\%$4 savings in generalized system costs relative to taxis, with average system cost scaling as $5\%$5 once maximum platoon size is reached (Fan et al., 26 Nov 2025).

The literature also converges on several objective cautions. Semi-on-demand is not synonymous with fully flexible routing: as demand and region size increase, optimal design can shift from fully-flexible to semi-flexible and finally to fixed-route service (Zhen et al., 2024). Higher demand does not automatically imply higher feeder efficiency when fleet size is fixed, especially in rural settings where circuitousness rises rapidly (Lee et al., 2019). Multimodal feeder–rail systems are not uniformly beneficial at low adoption: in Berlin and Brandenburg, the bi-modal system became advantageous only when a sufficient proportion of the population adopted it, with the study highlighting $5\%$6 adoption as the regime where energy and traffic reductions became substantial (Sharma et al., 2023). Finally, passenger gains and generalized-cost gains need not coincide: RL-controlled zonal SoD in Munich increased the number of served passengers, but did so at a higher average generalized cost than conventional fixed route (Ng et al., 2 Sep 2025). These results indicate that semi-on-demand feeder service is best treated as a design problem of regime selection, not as a universally dominant replacement for fixed-route transit.