Post Millennium Renaissance · Ground transport autonomy United States · 2026
Theme:

The car drives itself.
The train still needs a human.

Intuition says the train is the easy one — fixed rails, no pedestrians, no left turns into traffic. Yet in the United States, a robotaxi will carry you across a city with nobody in the seat, while almost every train you can ride still has a person in the cab. This is a snapshot of that inversion, why it exists, and what it would take to close it.

On public roads

Driverless cars

~500k
paid fully driverless rides per week, no human aboard, across 10+ US metros. Real, commercial, scaling fast.
On the national network

Driverless trains

~0%
of light rail, subway, commuter, intercity & freight run with no one in the cab. The lone exception is a closed-guideway system in Honolulu.
BART could drive itself in 1972. Waymo's founders weren't born yet.
Days since a US train was engineered to run driverless:
A

Every US train mode, plotted on the automation ladder

The rail industry grades automation GoA 0 → 4. GoA 4 means unattended: movement, stopping, and doors all run with no driver and no attendant aboard. The solid bar shows how far each US mode actually goes today. The marks where a human is still required in the cab. Tap any mode for the full story.

Show: Solid bars = what runs today.
Reaches driverless (GoA 4) Automation assists, crew aboard Essentially manual
GoA 0
manual
GoA 1 GoA 2 GoA 3 GoA 4
driverless
Human still legally required in the cab
Tip: tap a mode to expand it, or switch the toggle to compare today vs. what's technically possible.
B

So why is the "easy" problem the one that's stuck?

It isn't the technology. Driverless metro tech has existed for 40 years and runs on 60+ lines worldwide. The US gap is mostly law, labor structure, and where the track runs.

The real bottleneck

A 1964 labor clause, not a hard engineering limit

Federal transit funding carries Section 13(c), a worker-protection rule that makes it extraordinarily hard to remove existing rail jobs. So automation appears precisely where that rule doesn't reach.

BART was designed for driverless running when it opened in 1972; DC Metro had auto-operation capability from 1976. In both cases the onboard operator was a labor concession, not a technical necessity.

The trains could already drive themselves. We decided to keep a person there.
Why cars leapt ahead

Open road, light rules, deep capital

Robotaxis face no equivalent of 13(c) and no incumbent "driver's union" gating deployment. A single company put ~2,500 driverless vehicles on the road and crossed ~500k paid rides a week in under three years.

Trains carry a heavier safety-case burden too: an open freight right-of-way must perceive grade crossings, trespassers and obstacles for miles ahead — far harder than a closed metro guideway, which is exactly why Honolulu's sealed, elevated line could go fully driverless while a freight main can't.

The closed-vs-open distinction is the whole game: airport trams and Skyline run in environments nothing can wander into. The national network is the opposite.

1972
BART opens already engineered for driverless operation — operator added by agreement.
1985
Vancouver SkyTrain runs fully driverless — the North American template, just over the border.
2023
Honolulu Skyline opens as the first US mass-transit GoA 4 line — because it was a brand-new system with no legacy crew.
2025
Freight's first pilot: regulators clear a crewless battery short-haul test in Georgia. The mainline gap finally cracks.
C

Quiz: which is harder to automate?

Five real-world scenarios. For each one, pick which vehicle has the harder job handling it autonomously. Most people get at least one wrong — and the misses explain the whole paradox.

Scenario 1 of 5
A child chases a ball into the path of the vehicle. Which is harder to automate?
Scenario 2 of 5
The vehicle needs to stop for an obstacle it just detected. Which is harder?
Scenario 3 of 5
Navigating the route itself — deciding where to go and how to get there. Which is harder?
Scenario 4 of 5
A rural grade crossing at dusk: a stalled pickup is on the tracks / in the road ahead. Which is harder?
Scenario 5 of 5
Getting legal permission to remove the human. Which is harder?
↑↓

The Grades of Automation, defined

Rail doesn't grade automation by how smart the train is — it grades by who is responsible for each job. The international standard (IEC 62290-1, used by the UITP transit association) tracks four functions and asks, at each level, whether a human or the system owns them: setting the train in motion, stopping it, closing the doors, and handling things when something goes wrong. As you climb GoA 0 → 4, those four jobs move one by one from the person to the machine.

Who is responsible for…
GoA 0GoA 1GoA 2GoA 3GoA 4
Setting the train
in motion / speed
Human Human System System System
Stopping the train
Human Human System System System
Closing the doors
Human Human Human System System
Handling disruption
(faults, emergencies)
Human Human Human Attendant System
Can system override
a human mistake?
(ATP / emergency brake)
No Yes Yes Yes Yes
Staff on board?
Driver Driver Driver Attendant None
Human owns it On-board attendant (not a driver) Automated system owns it No protection — system cannot intervene

Each grade in plain terms

GoA 0On-sight

Manual driving — no signaling, no protection

A person drives entirely by what they can see — signals, obstacles, hazards — with no system watching over them and no automatic backstop. If the driver misses a stop signal or overspeeds, nothing intervenes. This is how a streetcar in mixed traffic runs. Speeds stay low because the human is the only safeguard.

This is the key distinction from GoA 1: at GoA 0 the system cannot override the driver. At GoA 1 it can — that single addition (Automatic Train Protection) is what separates on-sight from protected manual operation.

US example · most street-running light rail & trams
GoA 1Protected manual

Human drives, a safety system can intervene

The driver still controls the train, but an Automatic Train Protection layer can brake automatically if they pass a stop signal or overspeed. The US calls its version Positive Train Control (PTC). The human is in charge; the system is a backstop.

US example · Amtrak, commuter rail, most freight
GoA 2STO · semi-automatic

The train drives itself, a driver still rides along

Starting, accelerating, and stopping are automatic. A driver remains in the cab to close the doors and take over if something goes wrong. Metros have run this way for 40+ years. Several US subway lines are GoA-2 capable but keep the operator aboard.

US example · automated subway lines (operator retained)
GoA 3DTO · driverless

No driver — but a staff member is still on the train

There's nobody in a cab. An attendant rides on board to manage the doors, help passengers, and handle emergencies, but does not drive. The cab itself can be removed from the design. Rare in the US.

US example · effectively none in mainline service
GoA 4UTO · unattended

Fully automatic, with no staff aboard at all

Movement, stopping, doors, and disruption response are all handled by the system — no driver, no attendant. This only works in sealed environments nothing can wander into, which is why it runs on airport trams and Honolulu's elevated Skyline, but not on an open freight main.

US example · airport people-movers · Honolulu Skyline

Worth knowing: GoA 4 is not new or experimental. More than 70 metro lines across 40+ cities worldwide run unattended today, and the first driverless train experiments date to the 1920s. The US gap (see ) is about labor law and open track — not about whether the technology works.

🌐

USA · Canada · Switzerland — mode by mode

Same GoA scale, three different countries. The bars show today's deployed reality, not ambition. The Why tag inside each cell is the honest reason the bar stops where it does. Click any country name to highlight its column.

Mode
🇺🇸 USA
🇨🇦 Canada
🇨🇭 Switzerland
Automated metro / guideway
Closed, grade-separated systems
GoA 4
✓ Honolulu Skyline (2023); airport people-movers for decades
GoA 4
Vancouver SkyTrain driverless since 1985 — world's longest fully automated rapid transit system
No closed automated metro system exists in Switzerland. Cities are served by tram and light rail
Light rail / tram
Urban surface rail
GoA 0–1
Mixed traffic = on-sight driving. Technology limit, not just policy
GoA 0–1
Street-running systems (Toronto streetcar, etc.) require on-sight operators. Same physics limit as USA
GoA 2
Waldenburgerbahn (Basel area) reached GoA 2 Jan 2026 — first in Switzerland. BLT now expanding CBTC to tram network. GoA 4 targeted by 2030
Urban subway / heavy metro
Existing underground lines
GoA 1–2*
GoA 2 capable; operator retained by Section 13(c) labor law. BART built driverless-ready in 1972
GoA 1 / GoA 4
TTC (Toronto) GoA 1 — union opposed automation since 1985
SkyTrain (Vancouver) GoA 4 since 1985 — no legacy operator agreement
No heavy metro. Swiss cities use tram + regional rail as primary urban transit
Commuter / regional rail
Metra, LIRR · GO Transit · SBB S-Bahn
GoA 1 + PTC
Driver + Positive Train Control safety layer. Section 13(c) limits change
GoA 1
GO Transit runs human-driven. ETCS being introduced on Kitchener Line (Metrolinx), foundation for future GoA 2
GoA 1 + ETCS
SBB S-Bahn runs on full ETCS Level 2 (complete by 2025). Foundation for GoA 2 ATO under smartrail 4.0 (rollout 2027–2038)
Intercity / long-distance rail
Amtrak · VIA Rail · SBB IC
GoA 1 + PTC
Human engineer, open mixed-use network. No automation roadmap on mainline
GoA 1
VIA Rail runs on CN/CP freight corridors — human crew, signalling similar to US. High-Frequency Rail corridor proposal in planning
GoA 1 + ETCS
SBB IC on ETCS Level 2. GoA 2 ATO over ETCS tested (Lausanne–Villeneuve 2018). Targeted for GoA 2 deployment under smartrail 4.0
Freight rail
Class I / CN, CP / SBB Cargo
GoA 1 + assist
Driver-assist + PTC. First crewless pilot (Parallel Systems, Georgia) underway
GoA 1
CN and CPKC run human crews network-wide. Remote-control yard shunting is in use. No mainline crewless pilot
GoA 1
SBB Cargo runs human crews. Remote-controlled shunter trials conducted 2023–24 at Zürich-Mülligen yard
GoA 4 deployed GoA 2 deployed GoA 1 (manual + safety system) No equivalent mode ★ Leads in this category

Why each country is where it is

🇺🇸

United States — the legal bottleneck

The US has the technology and in some cases built systems that were ready for it in 1972. The constraint is Section 13(c) of the Federal Transit Act: worker protections that make removing rail jobs from federally funded projects legally very difficult. Automation happens only where that clause doesn't apply — new-builds like Honolulu and airport people-movers. The subway systems most ready to automate (BART, DC Metro) are the ones most locked by existing labor agreements.

1GoA 4 mass-transit line (Honolulu, 2023)
~0%of mainline, commuter, or freight running driverless
50 yrssince BART was built driverless-capable (1972)
🇨🇦

Canada — the same country, two opposite stories

Canada is the clearest proof that the bottleneck is institutional, not geographic or technical. Vancouver's SkyTrain has run fully driverless since 1985 — 40 years, 80+ km, 427,000 rides a day — because it was a greenfield system with no legacy workforce to negotiate with. Toronto's TTC tells the opposite story: the Scarborough RT literally drove itself but kept a human in the cab for union reasons, exactly mirroring the US. The Ontario Line (2031) will be GoA 4 because it's another new-start. VIA Rail and CN/CP freight mirror the US — human crews, GoA 1.

80+ kmGoA 4 driverless transit (SkyTrain) — world's longest automated rapid transit system
1985Year SkyTrain started — 38 years before Honolulu
2031Ontario Line opens GoA 4 in Toronto — another new-start unlocking what the TTC couldn't
🇨🇭

Switzerland — the network-first, whole-system approach

Switzerland has no Section 13(c) equivalent and no closed metro to start with, but it's taking a fundamentally different path: automate the whole network systematically rather than cherry-picking closed-system wins. SBB's smartrail 4.0 programme targets GoA 4 across intercity, regional, and freight by 2027–2038 — meaning even the mainline Bern–Zurich IC is in scope. The constraint isn't labor law; it's the genuine engineering challenge of open, mixed-traffic mainline track where trains share corridors with freight, cross roads at grade, and run across national borders. GoA 4 on an open mainline requires perception systems that don't yet exist at commercial scale. Switzerland is building toward that, not waiting for a shortcut.

GoA 2First Swiss railway to reach it: Waldenburgerbahn, January 2026
100%of SBB mainline on ETCS Level 2 — the digital foundation for GoA 2 ATO
2027–38smartrail 4.0 rollout targets GoA 4 network-wide — the most ambitious national programme of the three
The lesson across all three

Two different barriers, one outcome

The US and Canada share the same institutional barrier: legacy labor agreements on existing systems, and legal protection that freezes that state in place. Both countries achieve GoA 4 only on new-builds where that protection doesn't apply — Honolulu, SkyTrain, the future Ontario Line.

Switzerland faces a different, genuinely technical barrier: open mainline track is harder to automate than a sealed metro. But Switzerland is attacking the harder problem systematically, and has no equivalent of Section 13(c) preventing it from retrofitting existing systems once the technology is ready.

Vancouver proved in 1985 what the US won't let its subway systems become. Switzerland is building what neither North American country has attempted: a whole-network automation roadmap that includes the mainline.

Sources for this tab: SkyTrain Wikipedia (ridership, 1985 opening, world's longest); Alstom press release July 2025 (Mark V trains); urban-transport-magazine.com / railwaypro.com (Waldenburgerbahn GoA 2, Jan 2026); railtech.com (smartrail 4.0, GoA 4 target 2027–38); railwaygazette.com (Waldenburgerbahn GoA 4 target 2030); railjournal.com (SBB no plans for driverless mainline); Wikipedia Toronto subway (TTC union opposition, Scarborough RT driver retention); toronto.ca / metrolinx.com (Ontario Line GoA 4, ~2031); ifp.org (Section 13(c) analysis, Vancouver precedent). Confidence tier: GoA levels for deployed systems = Verified. Switzerland smartrail 4.0 targets and Canada GO Transit ETCS = Modeled from programme documentation.

A

Where the money goes — US transit operating costs

Labor is the dominant cost in running a transit system. This is the budget context that makes automation a financial argument, not just a technology argument. When 59 cents of every operating dollar pays wages and benefits, removing a cab operator is a structural budget conversation — not a gadget upgrade.

59%
of US transit operating costs go to labor — wages, benefits, and paid absences — according to APTA's 2025 data from the National Transit Database. Automation's primary financial case is here. Paris cut operating costs by 30% when it converted Metro Line 1 to GoA 4. One study estimated automation could cut US rail transit operating costs by up to 46%.

Source: American Public Transportation Association (APTA) via IFP.org analysis of FTA National Transit Database 2025. Remaining categories estimated from FTA NTD operating expense object class breakdowns (vehicle maintenance ~15%, non-vehicle maintenance ~8%, general admin ~10%, purchased transport ~8%). Tier: Modeled — labor % is Verified; sub-category splits are approximate from NTD object classes.

B

Canada's internal divide — rapid transit km by automation level

Within a single country, same decade, nearly identical technology — two cities made opposite choices. Vancouver built driverless. Toronto kept a human in the cab. The chart shows what that means in track kilometers operating today.

80 km
of fully driverless rapid transit in Canada — Vancouver's SkyTrain, GoA 4 since 1985, now the world's longest automated rapid transit system. Toronto's TTC subway runs 99 km of heavy metro at GoA 1 — same era, same continent, human in the cab — because the Scarborough RT's operators were retained by labor agreement when it opened in 1985, and that precedent held. The Ontario Line (~15 km, GoA 4) opens ~2031.

Sources: SkyTrain Wikipedia (79.6 km, GoA 4); Toronto subway Wikipedia (99.4 km total system, GoA 1 heavy rail lines). Note: Toronto's Line 5 Eglinton (opened 2026, GoA 2 underground section) is excluded as a light rail line for comparability with SkyTrain's rapid transit classification. Tier: Verified.

C

Urban metro km by GoA — 16 countries compared

Same GoA 0–4 scale, applied to urban metro systems only — no mainline, no freight, no commuter rail. This is the cleanest cross-country comparison possible: sealed, grade-separated passenger systems where the technology is proven and the only question is how far each country has deployed it. Each bar represents total urban metro km, split by automation grade. Sort and toggle views below.

GoA 4 — fully driverless GoA 2/3 — semi-auto, driver/attendant aboard GoA 0/1 — manual / protected manual
Sort by:
176 km
Riyadh, Saudi Arabia — world's longest GoA 4 metro, 100% driverless, opened 2024–25
~100 km
Singapore — NEL, DTL, TEL fully GoA 4; NSL and EWL upgraded to GoA 2 CBTC
~5 km
USA — Honolulu Skyline only GoA 4 mass-transit line; ~60 km GoA 2 (CBTC lines in NYC, DC, SF)

Sources & confidence: Riyadh 176 km — Gulf News / Guinness World Records 2025 (Verified). Dubai 101 km (Red/Green/Tram) — Railway Technology (Verified). Copenhagen Metro 43 km GoA4 — Wikipedia, Grokipedia (Verified). Sydney Metro 52 km GoA4 — Wikipedia / Sydney Metro official (Verified). Canada 80 km GoA4 (SkyTrain) + 99 km GoA1 (TTC) — Wikipedia (Verified). Singapore NEL/DTL/TEL ~100 km GoA4, NSL/EWL/CCL ~140 km GoA2 — Wikipedia, SGTrains (Modeled — line km summed from individual line articles). France Lines 1,4,14 ~61 km GoA4, Grand Paris Express phase 1 partial; remaining ~185 km mix of GoA1/2 — Paris Metro Wikipedia + railwaypro.com (Modeled). Germany Nuremberg U2/U3 ~19 km GoA4; Berlin/Munich GoA2 ~30 km estimated; rest GoA1 (Modeled). Japan Toei Oedo ~40 km GoA4; Tokyo Metro ~190 km GoA2 (all lines have ATO); some GoA1 — Grokipedia/Wikipedia (Modeled — Oedo alone confirmed GoA4; Metro GoA2 extent estimated). UK DLR ~38 km GoA4, Jubilee/Victoria/Elizabeth partial GoA2-3 ~40 km estimated; rest GoA1 (Modeled). USA Honolulu ~5 km GoA4; CBTC lines (L/7/Queens Blvd + DC ATO + SF partial) ~60 km GoA2; rest GoA1 (Modeled — GoA2 share is approximate). Switzerland Waldenburgerbahn ~13 km GoA2; everything else GoA1 (Verified). Qatar 76 km GoA4 (Doha Metro, all 3 lines) — Wikipedia / Doha Metro official (Verified). Spain Barcelona L9/L10 ~30 km GoA4 (open sections); Line 11 GoA3 ~2.3 km; rest GoA1 — Barcelona Metro Wikipedia (Modeled — central L9/L10 section not yet open). Mexico Lines 1 (18.6 km) & 12 (~18 km) CBTC GoA2 (Siemens/Alstom); total network ~226 km — railway-technology.com Dec 2025 (Modeled — GoA2 line count estimated from confirmed CBTC deployments). China ~900 km GoA4 (Shanghai Lines 10/14/15/18/Pujiang ~169 km confirmed GoA4; Beijing, Guangzhou, Shenzhen, Chengdu and others add ~730 km estimated); ~500 km GoA2; remainder GoA1 across 10,976 km total — Shanghai Metro Wikipedia, Grokipedia urban rail China, Railway Gazette China metro report Mar 2025 (Modeled — national GoA4 total is a fast-moving estimate; China adds hundreds of km per year).
Important scope note: Urban metro only. Mainline, commuter, freight, and light-rail-in-mixed-traffic excluded. Total km figures include only closed, grade-separated rapid transit.

C

Six major US systems — how far each has actually gone

The national average hides a wide spread. A few systems are GoA 2 on most lines. Some major cities are running 100-year-old signal technology. And the Philadelphia area — quietly — is home to the first ATO line ever built in the United States, still running today. Click any system card for the full story.

Note on SF trolleys / cable cars: San Francisco's iconic cable cars and historic streetcars operate at GoA 0 — manually operated, no automation, by deliberate design. MUNI Metro's light rail is GoA 1. They're excluded from the grid above as a separate category from rapid transit.

Confidence & sources

Tiered honestly. "Verified" = directly stated in named, dated sources. "Modeled" = my characterization of a range or a category-level claim, not a single reported figure.

Verified Cars: ~500k driverless rides/week, ~2,500 vehicles
Waymo crossed roughly 500,000 paid, fully driverless rides per week (early 2026 reporting), operating ~2,500 robotaxis across 10+ US metros, with 200M+ autonomous miles logged. These are no-human-aboard trips. Tesla's Austin service still carries a safety monitor and is excluded from the "driverless" count.
· thedriverlessdigest.com — Waymo 2025/2026 stats
· cnbc.com — 450k weekly paid rides (Dec 2025)
· carboncredits.com — fleet size, miles
Verified Honolulu Skyline is the only US mass-transit GoA 4 line
Hitachi Rail and multiple outlets confirm Skyline (opened 2023) as the first true mass-transit GoA 4 (unattended) system in the US, distinct from airport people-movers. It runs on a sealed, elevated guideway with platform screen doors — a closed environment, which is what makes GoA 4 feasible.
· hitachi.com / railwayage.com / masstransitmag.com — Skyline opening
· railwaynews.net — GoA 4 / technical profile
Verified The bottleneck is Section 13(c), not technology
BART (1972) was designed for driverless operation and DC Metro (1976) had auto-train-operation capability; onboard operators were labor concessions. Automation proceeds where federal transit labor protection (Section 13(c)) doesn't apply — new-starts like Skyline and FAA-funded airport movers. GoA 4 is mature, running on 60+ lines worldwide since SkyTrain in 1985.
· ifp.org — "Close America's Transit Automation Gap" (13(c) analysis, BART/WMATA design history)
Verified GoA definitions & responsibility matrix
The five Grades of Automation and the four responsibility functions (motion, stopping, doors, disruption) follow international standard IEC 62290-1 as defined by the UITP. GoA 2 = STO (driver aboard, auto drive/stop); GoA 3 = DTO (no driver, attendant aboard); GoA 4 = UTO (no staff aboard). 70+ metro lines in 40+ cities run GoA 4; first driverless experiments date to the 1920s. The matrix's human/system/attendant split is taken directly from these definitions.
· railjournal.com — UITP/GoA five-grade breakdown, Thameslink GoA 2
· questglobal.com — IEC 62290-1 function allocation
· en.wikipedia.org / rail-kn.com — STO / DTO / UTO modes, 1920s history
Modeled "~0%" and the per-mode ladder positions
No public dataset reports a clean "% of US trains running driverless" by mode — because for mainline rail the answer is effectively zero, so the bars encode automation grade (GoA), not a fleet share. Light rail / commuter / Amtrak sit at GoA 1 with a required operator. Subways are GoA 2-capable (auto operation on some lines) but keep an operator aboard. Freight is "driver-assist": PTC plus energy-management and braking systems, crew required, with one crewless short-haul pilot (Parallel Systems, cleared by FRA, Georgia). Bar widths are my proportional rendering of these grades, not measured percentages — treat them as a map of "how far each mode goes," not exact values.
· railwayage.com — freight automation / incremental approach, Wabtec, MxV Rail
· mordorintelligence.com — North American lines at GoA 1; freight "consist management"
· railroads.dot.gov (FRA) — ATO sensor-platform program