Automatic Train Operation (ATO) delivers where the conditions are right for it, argues guest author Alberto Mandler. The harder question, and one the industry rarely deals with, is where exactly those conditions hold — and where they don’t.
About the Author
Alberto Mandler is CEO and Co-founder of DirecTrainS, an Israeli deep-tech company developing operational concepts for dynamic train formation on standard rail infrastructure.
This article is part of a series of guest-authored opinion articles. By publishing them, RailFreight.com hopes to engage the rail freight industry in a discussion on how to move the sector forward.
The trade press has largely made up its mind about Automatic Train Operation. It considers it to be the next layer of rail digitalisation, a natural extension of ETCS, and a step toward higher capacity and lower operating costs. The capability arguments are real, since ATO does what humans cannot reliably do at scale. It manages train motion with millisecond-level precision, holds tighter headways than a fatigued driver, and accelerates and brakes with a consistency that smooths out flows across the network.
Where ATO is deployed in the right context
This rule applies everywhere. The gains are visible and measurable. London’s Thameslink core operates with ATO over ETCS Level 2 to target 24 trains per hour through one of the world’s most constrained junctions. The Elizabeth Line uses ATO to manage central-section throughput at frequencies that would be impossible under driver-controlled operation.
Rio Tinto’s AutoHaul system in the Pilbara runs unmanned heavy-haul freight trains across thousands of kilometers of dedicated infrastructure. Driverless metro lines operate in dozens of cities, with safety records that match or exceed those of driver-operated systems. In all these contexts, ATO delivers on its promises.
What the consensus skips over is the common thread running through those deployments. Every one of them operates in an environment where the constraints line up with what ATO actually optimises. The traffic is homogeneous, the operations are repetitive, the signalling regime is fixed, and the interaction with the wider network is limited. In these contexts, the binding constraints on capacity are driver consistency and minimum headway, which are precisely what ATO targets.
The real test comes when the industry pushes the same technology into environments with different constraints.
Where ATO’s constraint binds, and where it doesn’t
Most of the world’s mainline rail does not operate under the conditions in which ATO was first proved out. European mainline rail networks integrate high-speed passenger services with regional stopping trains and heavy freight on shared infrastructure. This is characterised by varied stopping patterns, diverse acceleration profiles, and separate signalling certification regimes.
In the United States, Class I freight corridors accommodate long-haul freight of varying lengths and weights, interspersed with Amtrak and regional passenger services, spanning hundreds of miles of track with varying grades and curvature. These conditions are not the environment for which Automatic Train Operation (ATO) was originally designed.
Between countries and regulatory regimes, capacity constraints vary. On a diverse mainline, the primary limitations are different. The first is network topology: junctions, merge points where multiple corridors meet, and single-track sections that reduce capacity due to directional changes.
The second is the timetable structure and the challenge of scheduling fast and slow services on the same corridor without the slow service occupying the time needed by the fast one. The third involves heterogeneity: a freight train gaining speed from a yard and a passenger service both requiring the same time slot on the same track. Finally, coordination across boundaries between operators, countries, and regulatory regimes is essential.
All of these are network-level problems, not driver-level ones. No train-control automation, however sophisticated, can reach them.
The headlines don’t offer the full picture
Extending ATO into that environment does not make the marginal gains vanish. A train under ATO control on a heterogeneous mainline still brakes more smoothly, maintains headway slightly better, and uses somewhat less energy. These gains are genuine, but they are limited by a network whose main inefficiencies lie at a different level.
The figures usually wheeled out to justify wider deployment carry less weight than they first appear to. Studies on ATO over ETCS Level 2 often report capacity increases of 15 to 25%, with some vendor marketing claiming up to 30%. However, these figures are mostly derived from simulation studies on dedicated corridors with uniform fleets and optimised timetables. Their conditions differ significantly from the heterogeneous mainline environments for which widespread ATO deployment is intended.
Independent academic research presents a more cautious outlook. In mixed-fleet operations, the proportion of fitted-to-unfitted trains has little impact on punctuality or capacity utilisation. Instead, capacity improvements primarily depend on station and junction geometry, not on inter-station block occupation, which is what ATO optimises.
A sober look at Automatic Train Operation
Furthermore, the operational principles of ATO over ETCS on mixed-traffic mainlines have not yet been fully defined, which itself makes the capacity estimates uncertain, and in every case the achievable gain depends on the specific operating environment, simulated corridor by corridor, so the headline numbers do not transfer from one deployment to the next.
Removing the simulation-favourable assumptions shows that the realistic capacity gain on a typical heterogeneous mainline corridor is much lower than the headline figures. In scenarios dominated by heterogeneity, the gain approaches zero. This must be balanced against the ongoing costs of maintaining an extra control layer. An added layer includes onboard equipment installation, integration across various train types, regulatory certification, lifecycle maintenance, and continuous coordination with drivers who oversee unautomated segments.
Aviation drew the same boundary decades ago
This is not the first time the question of ATO has been raised, nor is it unique to the rail industry. The aviation industry has tackled a similar challenge for many decades, and its solutions provide useful insights. Modern commercial autopilot systems are technically capable of operating across nearly every phase of flight, including taxi, takeoff, and landing under specific conditions.
The aviation industry has not chosen to deploy them universally. Autoland is certified and used in low-visibility conditions, but most landings under normal conditions remain pilot-flown. Taxiing is almost universally manual. Autopilot is engaged in cruise and during specific approach phases, where the constraints align with what the automation actually does. That means managing aircraft state across long stretches of relatively stable flight, with sparse traffic interaction and predictable boundary conditions.
The aviation industry did not extend autopilot into the phases where pilot judgment, real-time coordination with air traffic control, and reaction to non-routine conditions are the binding constraints. It’s not that the technology can’t do it. Rather, the design discipline advises against it, and that difference is the key issue.
What aviation engineered into its automation strategy, partly by design and partly through decades of incremental decisions, is a principle: a new control layer is only justified inside the deployment domain where its constraints align with the binding constraint at that scope. Outside that domain, the layer adds complexity without proportional benefit.
This is the principle the rail industry’s ATO discussion has been quietly missing. The technology is sound. The question is where and when it provides real gains, and what should be its scope of deployment.
Designing Automatic Train Operation for the segments where it truly pays off
The best solution is not to halt ATO development or to walk away from mainline deployment. It is to design ATO for the segments and zones where its constraints actually add capacityand to accept that those segments are finite subsets of the rail network rather than the whole.
These zones include yard outbound corridors, characterised by homogeneous traffic and repetitive patterns; junction approaches, where precise headways are crucial and geometry remains fixed; dense urban sections, where stopping patterns are predictable and throughput limits the system more than coordination; and interfaces between freight terminals and the mainline, where operational logic is clear and the broader network’s heterogeneity has not yet affected the process. These are the areas where ATO performs its best, intended function.
Across the rest of the heterogeneous mainline, the driver stays in charge — handling the coordination, the irregular conditions, and the network-level judgment that automation cannot reliably reach. The real engineering problemand a tractable one, is the handover between the automated zones and manual control. Getting the driver re-engaged cleanly, having the system make its own state plain across the boundary, and writing an operational protocol that lets ATO do what it is good at inside its own domain without taking on work drivers do better.
This is a narrower and quicker-to-deploy version of ATO than the universal-mainline ambition, and a more defensible one. It captures the capacity gains ATO does provide, in the segments where they are clear and measurable, and spares the industry the integration overhead of asking a train-level tool to do a network-level job.
The same logic applies to every rail control layer
None of this is unique to ATO. Every new control layer rail has taken on in the past three decades has hit the same problem in roughly the same shape. Positive Train Control in the United States proved its effectiveness gradually: primarily in collision avoidance, overspeed prevention, and work-zone protection. Once established as a routine layer of operations, it also took on additional coordination responsibilities.
ETCS Level 2 offers significant capacity improvements on busy passenger routes, where the primary constraint is moving-block headway. However, on freight-dominant corridors, where scheduling constraints are more limiting than headway constraints, its contribution to capacity is comparatively lower. Each technological layer functions best within its native domain, serving as the appropriate solution in that context, and can be extended beyond it when and where its benefits justify the effort.
This rule applies everywhere. Before deploying a new control layer, the first thing to determine is which constraint it targets and at what level. The scope of the deployment should be based solely on that. Significant improvements in rail come from layers that each focus on their own constraint and do their specific job. They do not come from stretching one layer to handle every task at all scales.
We need a disciplined, systemic vision
Aviation has successfully integrated this deployment discipline into its automation strategy over many decades through incremental boundary decisions. Rail can proactively establish this discipline now, before future control-layer expansions complicate the decision-making processes.
Rail’s capacity problem is real and urgent, and there are plenty of technologies to solve it. However, the industry lacks a disciplined, systemic vision for deploying each technology where it is most effective.
Is the rail industry ready to move beyond questioning the deployment of ATO to identifying where its limitations truly occur? And what other innovations are necessary in areas where those constraints apply?
