AMR fleet sizing is a capacity-demand balance, not a vendor catalogue count: the number of robots equals peak trip demand multiplied by average cycle time, divided by productive hours per robot, then uplifted for charge, traffic and contingency margin.
The two mainstream approaches in 2026 are vendor fleet calculators (KUKA, KINEXON) for a fast pre-feasibility figure, and closed-queue or mixed-integer optimisation models for a defensible production-grade number, with both routes converging on the same core equation but differing on charge strategy and assignment policy.
The core fleet-size equation and what each term really means
Minimum fleet size N is set by N = (R x T) / (H x U), where R is the peak trip rate (transports per hour the lines demand), T is the average round-trip cycle time including pick, travel, dropoff and queue, H is the available productive hours per robot per shift, and U is the utilisation ceiling the operator is willing to run at [S4]. A 2026 study on AMR-fed manufacturing systems formalises this as a capacity-demand balance where, below a certain robot count, stations starve, and above a threshold, each additional robot contributes only marginal throughput [S4].
Charging strategy, corridor sharing and battery-swap versus opportunity-charge policy each move this number by 10-25% in published models, which is why an OEM spreadsheet alone is not enough [S4].
Vendor calculator pass: KUKA and KINEXON inputs versus outputs
The KUKA AMR Fleet Calculator is a free, web-based pre-feasibility tool that takes plant inputs (transport routes, payload, shift model) and returns an indicative AMR count, AMR model selection, and a sizing summary intended as a starting point rather than a final bill of materials [S1]. KUKA explicitly flags the output as a guide and routes the user to its AMR experts for precise planning, which is the correct way to treat any vendor spreadsheet [S1].
The KINEXON AMR/AGV Fleet Size and ROI Calculator layers an ROI model on top of the count: enter the planned mobile-robot operation parameters and the tool returns the number of robots, the number of charging stations, and a payback estimate [S2]. KINEXON's positioning, vendor-agnostic orchestration of VDA 5050 AMRs and AGVs, matters here because fleets of more than 8-10 robots in 2026 deployments almost always need a fleet manager to handle traffic, charge and task allocation rather than running each AMR as an island [S2].
When the vendor number is wrong: the standalone-robot penalty
Standalone AMRs deployed without fleet orchestration deliver only 40-55% of their theoretical throughput, because robots collide for resources, idle at chargers and fail to coordinate with the production schedule, which means the calculator output must be marked up, not taken at face value [S3]. iFactory's 2026 field data also shows a 38% average reduction in internal material-handling cycle time after a managed AMR fleet is in place, with mid-size plants reporting $2.4M annual labour and downtime savings against a typical $1.8M-$4.2M baseline of internal-logistics waste [S3].
The implication for sizing is concrete: if the calculator says 6 robots, the engineering floor before contingency is 6 / 0.5 = 12 if those robots will run without a fleet manager, or 6 / 0.85 = 8 with orchestration. This is the single largest hidden multiplier in 2026 AMR projects, and the reason both AMR robot reference architectures and the calculator inputs have to be read together.
Decision criteria: rule-based, heuristic, or DRL assignment policy
Once the count is fixed, the assignment policy changes the effective capacity by 15-30% in published benchmarks, and the policy choice is therefore a sizing input, not a downstream detail [S4]. The 2026 MDPI survey on AMR-fed manufacturing identifies four common rule families: nearest-task, earliest-delivery-time, station-starvation-priority, and charge-threshold, with hybrid and data-driven policies such as deep reinforcement learning (DRL) outperforming classical rules on dynamic, multi-line plants [S4].
For a single-line, low-variance plant, a nearest-task or earliest-delivery-time rule is sufficient and the calculator's count holds. For multi-line plants with mixed takt times and shared corridors, specify DRL or at minimum a charge-threshold-aware heuristic, and add one robot to the calculator output per 10 lines of shared-corridor traffic, mirroring how concrete batching plant designers add trucks to a pour queue for truck-cycle variability. Charge strategy (battery swap, inductive, opportunity) and delivery time windows are treated together in the 2026 MILP literature as coupled constraints, not independent levers [S4].
Failure modes and constraints the spreadsheet does not show
Three failure modes dominate real 2026 deployments: starvation at the line when robot count is set too low, collision-and-idle when count is set too high without a fleet manager, and cascade delay when one robot failure propagates because no spare unit exists in the pool [S4]. Below the minimum-robot threshold the closed-queue models predict station starvation and takt loss; above a higher threshold, the marginal throughput gain per added robot collapses, which is the textbook S-curve and the basis for not over-buying [S4].
Operational constraints to add to the model: battery-swap versus opportunity-charge versus inductive-charge (each shifts the productive-hours-per-robot figure), corridor sharing between AMRs and forklifts, delivery time windows from the MES, and a contingency margin sized to the plant's mean-time-to-recover for a robot, normally one or two spare units per 20-robot pool. None of these are exposed in the KUKA or KINEXON calculators at the input level [S1][S2], which is why the OEM pass is step one and a queueing or MINLP validation is step two [S4].
Selection criteria and a four-option comparison for sizing method
For a 2026 plant project, the four practical sizing options line up against decision criteria as follows. Vendor spreadsheet only (KUKA or KINEXON): fast, free, indicative; weak on charge and failure; suitable for feasibility only [S1][S2]. Closed-queue analytical model with MINLP: defensible count, captures starvation and the marginal-robot threshold, but needs a specialist and assumes stationary demand [S4]. Discrete-event simulation (AnyLogic, FlexSim): handles shared corridors, charge strategy and failure injection; costlier to build; the right tool for plants with high takt variance [S4]. DRL or hybrid MILP-DRL: best for dynamic re-allocation on multi-line plants, highest performance, highest data and integration cost [S4].
Use the vendor tool for the first number, the closed-queue or MINLP model for the defensible number, and a discrete-event or DRL layer only when the plant has more than three coupled lines, shared-corridor traffic or aggressive variability targets, with the final spec written as a range, not a single integer, and a one-robot-per-20 contingency written into the procurement line.
Standards, sourcing and what to pin in the spec
No single ISO or IEC standard prescribes an AMR fleet count; the relevant 2026 references are VDA 5050 for the AMR-fleet-manager interface (cited in KINEXON's vendor-agnostic orchestration positioning) and the MDPI 2026 queuing-network and MINLP framework for the calculation method itself [S2][S4]. Pin the VDA 5050 conformance clause in any AMR purchase so the fleet manager is swappable, and pin the calculation method (closed-queue with MINLP, or DRL) in the system-engineering section so the count is auditable [S2][S4].
Trackable signals for the next planning cycle: published 2026 benchmarks from iFactory showing 38% cycle-time reduction and 96% fleet uptime with predictive maintenance against a 71% reactive baseline, and the MDPI 2026 model's finding that DRL-based assignment outperforms classical rules on dynamic multi-line plants [S3][S4]. A useful adjacent read on queue-based sizing logic applied to a different vehicle class is the mixer-truck fleet sizing method for a 60 m³/h pour, which uses the same capacity-demand balance against takt-time variability; the DRL-and-transformer AMR re-planning 2026 review is the right next read if the plant is heading toward a data-driven assignment policy.
Spec-level background on the components involved: pressure transmitter.