Robotics
Choosing Between AGVs and AMRs: A Decision Framework for Operations Leaders
The cost difference is real, the use cases overlap, and the wrong choice locks in years of operational rigidity. A structured decision framework grounded in throughput, layout, and TCO.
An autonomous mobile robot that navigates a clear aisle is a solved problem and has been for a decade. A fleet of forty that sustains contracted throughput through a shift change, a pallet left in a cross-aisle, and a localisation failure in a featureless corridor is not solved by better navigation. It is solved by traffic management, safety architecture and honest arithmetic about how many robots the work actually requires.
Fleet sizing: the calculation that decides whether the project succeeds
Nearly every disappointing AMR deployment we are asked to review shares a root cause, and it is not technical. The fleet was sized on nominal cycle time and nominal availability, and neither survives contact with a real facility. The correct treatment is Little's Law applied to the transport system, with availability derated for charging, congestion and fault recovery.
N = ⌈ ( λ × Tcycle ) / ( Au × ηc ) ⌉
The term teams omit is η<sub>c</sub>. Congestion efficiency is not a constant — it degrades non-linearly with fleet density, because robots do not merely slow each other down, they generate blocking events that cascade. Below roughly 0.6 robots per 100 m² of navigable floor we typically measure η<sub>c</sub> above 0.9. Past about 1.2 it falls below 0.7 and the marginal robot starts removing throughput from the fleet.
| Factor | Typical nominal | Realistic | Driver |
|---|---|---|---|
| Uptime A_u | 0.95 | 0.82–0.88 | Charging, faults, manual recovery |
| Congestion η_c | 1.00 | 0.70–0.90 | Fleet density, aisle topology |
| Cycle time | Straight-line | +18–35% | Queueing at pick/drop, lift waits |
| Handover dwell | Ignored | 20–90 s | Human or conveyor interaction |
| Peak : mean ratio | 1.0 | 1.4–2.2 | Shift patterns, order waves |
Localisation degeneracy, and why long corridors break scan matching
LiDAR localisation estimates pose by aligning the current scan against a prior map. In a geometrically rich environment this is well-conditioned. In a long, straight, featureless corridor it is not: the scan constrains lateral position and heading strongly, but provides almost no information along the corridor axis. The optimisation problem becomes ill-conditioned in that direction and the pose estimate drifts freely along it.
This is observable before it becomes a failure. The Fisher information matrix of the scan-match problem has an eigenvalue that collapses along the degenerate direction, and its condition number is a usable online health metric.
κ(H) = λmax(H) / λmin(H) , H = J⊤J
The engineering response is layered rather than clever. Monitor the condition number continuously. Where degeneracy is structural — long corridors, large open floors, aisles whose contents change daily — add observability the geometry does not provide: retroreflective markers at known survey points, ceiling fiducials, or floor-embedded RFID at decision points. Fuse wheel odometry and IMU so the estimate degrades gracefully rather than jumping. And crucially, publish covariance honestly, because a planner that trusts an overconfident pose is a planner that will drive into a rack.
Multi-robot path finding is where the real difficulty lives
Single-robot planning is a well-understood two-tier problem: a global search over a static costmap and a local controller that tracks the plan while avoiding what the map does not contain. Fleet coordination is a different problem class. Optimal multi-agent path finding is NP-hard, so production systems do not solve it optimally — they solve it well enough, fast enough, with guarantees about what cannot happen.
- Prioritised planning: order the robots, plan each against its predecessors' reserved space-time. Fast and simple, but incomplete — it can fail to find a solution that exists.
- Conflict-based search: plan independently, detect conflicts, branch with added constraints. Optimal, but its cost grows sharply with conflict density, so it needs a time budget and a fallback.
- Space-time reservation: robots reserve corridor segments for time windows. Deadlock-free by construction if reservations are acquired in a global order — the classic resource-ordering result applied to floor topology.
Whichever is chosen, deadlock freedom must be a property of the design, not something tested for. Four robots meeting at a cross-aisle, each waiting on the next, is a circular wait, and the standard prevention is to impose a total order on resource acquisition so a cycle cannot form. Detection-and-recovery is the weaker alternative and always ends in an operator walking to the aisle.
Safety is a separate architecture, not a software feature
Navigation obstacle avoidance and functional safety are different systems with different integrity requirements, and conflating them is the most consequential mistake in this domain. Costmap-based avoidance runs on a general-purpose computer executing non-certified code. It cannot be the reason a person is not injured.
Driverless industrial trucks are governed by ISO 3691-4, with safety functions designed to ISO 13849-1 performance levels and protective devices to IEC 61496. The practical shape of a compliant system is a safety-rated laser scanner driving a certified safety controller that can command a stop independently of the navigation stack — the application requests motion, and the safety layer holds an absolute veto.
Protective field geometry follows from stopping distance, which is speed-dependent, and this is why zone switching by velocity is mandatory rather than an optimisation.
dstop = v · ( tdetect + tresponse ) + v2 / (2a) + Zm
Two terms are routinely underestimated. Deceleration a must be the guaranteed value for a fully loaded robot on the most slippery floor condition in the specification, not the value measured empty on clean epoxy. And brake performance degrades over service life, so Z<sub>m</sub> must cover the maintenance interval, not the commissioning day.
Interoperability: assume you will not stay single-vendor
Most facilities that succeed with AMRs eventually run more than one make, whether by acquisition, by expansion, or because the second application had different requirements. VDA 5050 exists for this reason — a common MQTT/JSON interface between fleet manager and vehicle covering order assignment, node-and-edge graph execution, state reporting and actions.
It is worth adopting even in a single-vendor deployment, because the alternative is a proprietary integration that becomes the reason you cannot change vendor later. The cost of conforming is measured in weeks. The cost of not conforming is measured in the negotiating position you no longer have.
What separates a fleet that works
Honest sizing arithmetic with derated availability. Localisation that measures its own conditioning and degrades predictably. Coordination that is deadlock-free by construction. A safety layer that owes nothing to the navigation stack. And an interface standard adopted before it is needed. None of it is exotic. All of it is the difference between a fleet that holds contracted throughput in year three and a pilot that quietly stops being mentioned.
Explore how Root Digit can support your team
From discovery workshops to production deployment, our engineers and consultants partner with you across the lifecycle of your AI, robotics, and IoT initiatives.