Skip to content

About

Microscopic traffic-intersection simulator in Rust. Compare fixed-time, actuated, max-pressure, predictive (MPC) and deep-RL signal control under realistic sensors, pedestrians, buses and corridors, and see the delay each person experiences. Runs in your browser via WebAssembly or on the desktop.

Topics

Resources

Code of conduct

Contributing

Security policy

Accessibility

Stars

0 stars

Watchers

0 watching

Forks

Latest commit

 

History

99 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Intersection Lab

OpenSSF Best Practices OpenSSF Scorecard

A graphically rich, microscopic traffic-intersection simulator in Rust for exploring how signal control, sensing, pedestrian treatments and AI affect what drivers, transit riders and pedestrians actually experience, at one intersection or along a corridor. It also looks ahead to a signal-free crossing for fully automated, connected vehicles, to measure how much such a crossing could carry, why its vehicles should still take turns in batches rather than weave through one another, and what happens when one of them fails in the middle of it.

Try it in your browser, no install needed, or download it for your desktop below.

Arterial intersection under predictive control

Quick start

  1. Open the browser version, or install the app and start it. Traffic starts flowing right away.
  2. In the Scenario panel on the left, pick an Intersection and how it's controlled: a Control mode and, for traffic signals, a Controller. Watch the delay and queue charts on the right respond.
  3. Under Benchmark on the right, press ▶ Compare controllers to run every option on identical arrivals and chart the results.
  4. Click a vehicle to see its speed, delay, stops and who it's yielding to. Space pauses.

What you can do in the app covers the rest.

Install

Download the package for your platform from the Releases page: a universal DMG for macOS, an MSI (or portable zip) for Windows on x64 and ARM64, and .deb, .rpm or .tar.gz for Linux on x86_64 and aarch64. All are built by GitHub Actions and come with a signed build-provenance attestation, and the macOS app and the Windows executable and installer are also code-signed; see packaging/README.md for how releases are made and verified.

If you have a current Rust toolchain, you can instead build and install it from crates.io:

cargo install --locked intersection-lab

Run the same command again to update it, and cargo uninstall intersection-lab to remove it.

Build from source

cargo run --release                               # the interactive app
cargo run --release --bin bench                   # headless controller comparison (all layouts)
cargo run --release --bin bench sensors all       # sensor study: packages, weather, failed loops
cargo run --release --bin train all 800 900 dqn   # train the Deep-RL agents (writes weights/)
cargo run --release --bin train all 800 900 hybrid  # train the hybrid agents (writes weights/)
cargo run --release --example corridor            # corridor study (green wave, adaptive, upstream link)
cargo run --release --example lpi                 # pedestrian safety treatments
cargo run --release --example fya                 # left-turn phasing modes
cargo run --release --example satflow             # calibration against HCM saturation flow
cargo run --release --example cooperative         # capacity of the signal-free cooperative crossing
cargo run --release --example failures            # a vehicle fails there: what each gap survives
cargo test --release                              # regression tests: safety, throughput, app, tools

The browser version is the same app compiled to WebAssembly. Build and serve it locally with Trunk (rustup target add wasm32-unknown-unknown first); GitHub Actions publishes it to GitHub Pages from main (.github/workflows/pages.yml):

trunk serve --release --open                      # http://127.0.0.1:8080

In the browser, benchmarks run one strategy at a time instead of in parallel, and trained weights can't be saved; everything else matches the desktop app.

What you can do in the app

  • Intersections: 4-way arterial with left-turn bays, 4-way major/minor, 4-way single-lane town intersection, T-intersection, single-lane roundabout, and a signal-free cooperative crossing for automated, connected vehicles.
  • Corridors: set signals along main road (1–5). Intersections sit one block (420 m) apart; traffic flows between them, queues spill back, and fixed-time signals can run a coordinated green wave. The panels show end-to-end trip time and stops, plus any single intersection (◀ ▶ or click a vehicle).
  • Side by side: Compare side by side runs a second copy with identical arrivals under another strategy (B), in a split view with an A/B table. On the cooperative crossing, B is another schedule or safety gap; by default it weaves vehicles through first come, first served.
  • Control: traffic signal (6 controllers, hot-swappable while traffic runs), all-way stop, two-way stop, roundabout yield, and on the cooperative crossing an intersection manager that schedules every vehicle (no signals, no sensors), in batches or first come, first served.
  • Demand: per-approach volume and turning shares, global multiplier, vehicle mix, pedestrian volume, and arrival pattern (Poisson, platoons, rush-hour peak, surges).
  • Spawn by hand: pick a vehicle type and intent (left/through/right), then click an approach; pick Pedestrian and click a crosswalk corner.
  • Sensors: loops, cameras, advance radar, connected vehicles, and in a corridor an upstream link. Field-realistic sensor errors (on by default) adds weather and lighting conditions, failed loop stations and turn-signal use.
  • Operations: emergency preemption, right turn on red, a "no turn on red while pedestrians are present" blank-out sign, leading pedestrian interval, left-turn phasing (flashing yellow arrow / protected only / automatic), and a max-wait guard.
  • Failures (cooperative crossing): make a vehicle stop dead or brake to a stop, by clicking it, in the middle of the box, at the worst moment for crossing traffic or at random, and see who stops short, who doesn't, and how the manager keeps the open lanes moving until the vehicle is cleared away. Side by side, both strategies get the same failure, except at the worst moment, which each reaches on its own.
  • Inspect and measure: click a vehicle for its speed, delay, stops, fuel, detecting sensor and who it yields to. KPIs, live charts, delay histogram, breakdowns by approach, movement and class, signal timing diagram, controller "brain", pedestrian near-misses.
  • Benchmark: Compare controllers and Sensor study run every option headless on identical arrivals and chart the results. On the cooperative crossing, Compare schedules runs batches under each of its three safety gaps, and first come, first served under fail-safe gaps.

Keyboard: Space pauses. Mouse wheel zooms, drag pans.

How the simulation works

Piece Model
Car following Intelligent Driver Model, per-class parameters, desired-speed spread, 0.6–1.2 s driver reaction when pulling away from a stop
Calibration A saturated through lane of cars discharges at 1,863 veh/h/ln with 2.2 s start-up lost time (HCM base: 1,900 and ~2 s); the default mix (8% heavy vehicles) gives 1,698, in line with HCM's heavy-vehicle factor (examples/satflow.rs)
Yellow Dilemma-zone behavior: drivers under ~2 s from the line (or facing hard braking) continue; the all-red covers them
Turning speed Curvature of each connector → comfortable lateral acceleration
Lanes Left-turn bays are 75 m storage lanes carved out of the median; left-turners arrive in the next lane and merge in, and wait behind the bay queue when it is full (bay spillback). Discretionary lane changes use MOBIL (safety criterion + politeness-weighted incentive), never across the solid line in the last 25 m
Right of way Paths are intersected into conflict points with measured zones. Vehicles compare priority (signal, stop order, circulating vs entering) and predicted occupancy windows, accept gaps for the whole maneuver, and commit once they can't stop. A commitment grants priority, not room: committed vehicles still stop for anyone stalled in the zone, and a vehicle waiting inside a zone holds it. Wait-for cycles are broken by ignoring vehicles that are themselves waiting, and gridlock is avoided by never pulling into a crossing behind a vehicle that is still creeping up to where it will wait, and by not waiting inside the box across a path whose traffic is only held by a red light. If two vehicles commit to a crossing in the same instant, the one that can still stop yields; followers keep seeing a long vehicle that turns off until its tail clears the split
Corridors Vehicles leaving an intersection enter the next one's approach; when it is full they wait on the exit road, and drivers don't enter the box without room to clear it
Signals Stage-based state machine with yellow and all-red, pedestrian WALK + flashing DON'T WALK, min green stretched for pedestrians, right turn on red, optional LPI
Pedestrians Push-button calls, walk in compatible stages, vehicles yield; at unsignalized crosswalks they assert right of way to vehicles that can stop comfortably
Stats Control delay = travel time − free-flow time on the vehicle's own route; person delay weights by occupancy
Safety Separating-axis overlap test between segmented vehicle bodies, in the intersection and within lanes. tests/safety.rs and examples/sweep.rs require zero overlaps, and tests/safety.rs replays busy seeds that used to gridlock

Arrivals come from a dedicated random stream, so every controller faces the same vehicles.

Signal controllers

All controllers see only the SensorView, built from their sensors rather than ground truth. Pedestrians and emergency vehicles are the exceptions (below).

  1. Fixed-time (Webster). Cycle and splits from critical flow ratios of historical demand; every green covers its pedestrian interval. In a corridor it can be coordinated: common cycle, main-road stage first, spare time to it, offsets chosen to maximize the two-way green band, force-offs, no late walk.
  2. Actuated (NEMA-style). Min green, passage-time extension, gap-out/max-out, phase skipping, transit-priority extension. Detector stations flagged faulty go on max recall.
  3. Max-Pressure (Varaiya 2013). Serve the stage with the largest upstream-minus-downstream queue pressure.
  4. Predictive, schedule-driven MPC (in the spirit of SURTRAC). Every second, predicts each detected vehicle's arrival, simulates candidate schedules over 90 s, and minimizes person delay with a fairness penalty on long waits.
  5. AI: Deep RL (Double DQN). A pure-Rust MLP with one hidden layer chooses the next stage from per-movement queues, predicted arrivals, waits, bus presence, stage and pedestrians, under a safety shield (max green, no holding or switching to empty stages, a walk for any pedestrian who has waited 40 s).
  6. AI: Hybrid (RL-tuned predictive). The predictive planner makes every signal decision; a DQN re-tunes it every 10 s (fairness weight, hold/switch bias, max green: 48 tunings). Training first searches for the layout's best static tuning, which becomes the fallback and the bar to clear; the network is kept only where it beats that bar.

Safety layers sit on top of every controller: emergency preemption (emergency-vehicle delay 22.1 s → 4.0 s, examples/preempt.rs) and a max-wait guard.

Sensors and their failure modes

Sensor What controllers get Realistic errors
Inductive loops Stop-bar presence, advance pulses, spot speed, input-output queue estimate Stations fail stuck-on/off; motorcycles missed ~35%. Diagnostics (max presence 3 min; no activity for 10 expected headways) flag failures, and the lane falls back to volume-based estimates / recall
Intersection camera Tracks within range with intent, class and individual waits Detection falls with distance and weather (night, rain, fog, snow shrink range and hit rate); trucks and buses hide vehicles behind them; intent only from drivers who signal (80%), others look like through traffic
Advance radar Long-range position and speed ±1.5 m / ±0.5 m/s noise, small miss rate in rain/snow
Connected vehicles Position, speed, route, class at any distance GPS noise (latency compensated by timestamps)
Upstream link (corridor) Vehicles the upstream signal has released toward us Intent unknown

Pedestrians and emergency vehicles reach the SensorView straight from ground truth, whichever sensors are on. Controllers use the push-button calls and the number of people waiting at each crosswalk, with the longest wait. Only preemption reads the emergency vehicles within 300 m of the stop line; controllers see them only through their sensors.

The signal-free cooperative crossing

The last intersection is a look at a possible future rather than at today's streets. It assumes every vehicle is automated, connected and fitted with crash avoidance. It has no signals and no roadside sensors, and pedestrians cross in underpasses, so they never stop traffic. An intersection manager schedules every vehicle that crosses the box. In each lane it considers only the first vehicle without a slot; when it grants one, it sends a complete trajectory, and the vehicle drives exactly that. Right turns leave on slip lanes before the box and never ask for a slot.

The design notes in docs/cooperative-intersections.md work through the idea from first principles: what limits a conflict zone, what vehicles and the manager would need to share, how to schedule for throughput, a reference design for a large crossing, the safety case, security and privacy, mixed traffic, and the wider ramifications. Their last section compares the estimates with what this simulator measured.

Piece Design
Geometry Each approach has two left-turn lanes, three through lanes and a right-turn slip lane around a corner island, so right turns never enter the box. Opposite left turns pass clear of each other and every lane has its own receiving lane: paths in the box only cross, never merge. Default demand: 4,500 veh/h per approach (18,000 in all), 20% turning left and 15% right
Reservations Each pair of crossing paths has a conflict zone on each path. The manager builds every trajectory with the simulator's own integrator, so it knows when each vehicle will be in each zone, and grants it only if it keeps the required gap to every trajectory already granted. A granted trajectory never changes, unless a vehicle fails
Phases By default the manager serves each movement in batches, in two rings with a barrier like an actuated signal, ending a batch when no vehicle can continue it (gap-out) or at its demand-weighted share of a 100 s maximum cycle (max-out). Left turns lag the through movements: a left turn can follow the opposing through traffic almost at once (0.3 s), while the reverse switch would cost about 6 s. The next batch starts as soon as its zones are free, with no yellow or all-red. Emergency vehicles are served out of turn
First come, first served The alternative schedule has no phases. Of the vehicles first in their lanes without a slot, the one that could cross the stop line soonest goes next, with the earliest trajectory that fits around those already granted. It is the familiar picture of vehicles weaving through one another, and it carries about half as much (below)
Safety gaps Fail-safe (default): enough time for a follower to stop short of the zone if the vehicle ahead stalls in it: two message latencies, the actuation delay and the time its braking distance takes at crossing speed (1.43 s for a car at 40 km/h). Mitigated: it could still slow to 25 km/h. Nominal: the two message latencies and the actuation delay only (0.5 s), sized for a vehicle ahead that brakes hard to a stop, not for one that stops dead
Monitors Independently of the plans, the simulator measures what vehicles actually did: the closest post-encroachment time (PET), zones entered while occupied, box entries without a reservation, and any step off the granted trajectory
Failures A vehicle can be made to fail: it stops dead, as after a collision, or brakes hard to a stop. Half a second later (two message latencies and the actuation delay, which every gap allows for) the manager reacts with new trajectories. A vehicle that can still stop short of the box loses its crossing and stops there. One that can't brakes as hard as it can if the failed vehicle is in a zone on its path, or follows the vehicle ahead until it rests behind it. Everyone whose path stays clear crosses as planned. Vehicles that can't stop in time touch: they stop where they touch and are disabled too, and the speed at which they closed is the measure of it. No new crossing is scheduled across a zone that a vehicle at rest holds, the other lanes are served as usual, and vehicles that stopped in the box are routed out once their way is clear. Disabled vehicles are cleared away after a minute by default

Why weaving one at a time carries less

The popular picture of automated vehicles at an intersection is a zipper: each vehicle adjusts its speed on the approach and slips through a gap in the crossing traffic, and nobody ever stops. Scheduled that way, first come, first served, the crossing does better only in light traffic, and it saturates at about half the demand that batches carry. Average delay with fail-safe gaps, the default vehicle mix and the same arrivals for both (3 seeds × 20 min):

Demand (veh/h) First come, first served Batches
2,000 0.4 s 0.6 s
4,000 1.2 s 1.8 s
6,000 3.1 s 3.0 s
8,000 7.4 s 5.1 s
10,000 21.8 s 7.7 s
12,000 102 s, queue growing 12.5 s
14,000 186 s, queue growing 17.0 s
18,000 (design) 293 s, queue growing 28.6 s
Capacity about 10,100 about 19,000

Three things cost the zipper its throughput:

  • Every switch pays a safety gap; a follower doesn't. A vehicle following another on the same path needs only its car-following gap, about 1.0 s at 40 km/h. A vehicle crossing behind one from a conflicting path must wait for the first to clear the zone, about 0.7 s, and then for the fail-safe gap, 1.43 s for a car: about 2.1 s in all. Served one vehicle at a time, alternating between two paths, a single conflict zone carries about 1,700 veh/h; in batches of 8 it carries about 3,100.
  • Each path crosses ten other lanes. A weaving vehicle needs a free window at every one of them, at times its speed fixes. Fitting it into the first set of windows that works leaves slivers of time at the other zones, too short for anyone else.
  • One vehicle waiting holds up its lane. Vehicles can't pass each other in a lane, so a vehicle whose windows don't line up delays everyone behind it, while a batch keeps every lane of a movement streaming together.

Shorter gaps help the zipper more than batches, since they make switching cheap. With nominal gaps (latency and actuation delay only) first come, first served reaches about 12,600 veh/h, 25% more, while batches gain only 3.5%, to 19,700; the zipper still carries about a third less. So the schedule that carries the most looks like a signal without its yellow and all-red: each movement goes in a batch, and the next starts the moment its zones are free. To watch the difference, pick the cooperative crossing and turn on Compare side by side: B weaves the same arrivals through first come, first served, slows its vehicles far out on the approaches and, at the design demand, soon has a queue waiting to enter the map.

Batches (A) and first come, first served (B) side by side at the design demand

What the safety gap is for

The gaps are sized for a failure that normal running never shows, so the app can cause one: under Failures, a vehicle can be made to stop dead or to brake hard to a stop. The hardest case for crossing traffic is a vehicle that stops dead just as it is about to clear a zone, with a crossing vehicle due right behind it at the smallest gap the policy allows. Stall one at the worst moment waits for exactly that.

A car stopped dead in the box, with fail-safe gaps (left) and nominal gaps (right)

Twelve seconds after a southbound car, ringed in red, stopped dead as it was about to clear the path of a platoon of left turns. Left, with fail-safe gaps, the platoon has stopped short of it. Right, with nominal gaps and the same arrivals, its first vehicle reached it. Westbound traffic, whose lanes stay clear, keeps crossing.

One failure in the box per run, 400 runs per row, in batches at 16,000 veh/h and first come, first served at 8,000 veh/h (examples/failures.rs; the results have every case):

A vehicle stops dead at the worst moment Batches: crossing contact Fastest Closest emergency stop First come, first served: crossing contact Fastest Closest emergency stop
Fail-safe gaps none 0.6 m short of its zone none 0.4 m inside it
Mitigated gaps 85% of failures 21 km/h 4.5 m inside 54% 22 km/h 6.6 m inside
Nominal gaps 96% 35 km/h 12.5 m inside 89% 35 km/h 12.8 m inside
  • Each gap survives what it is sized for, with little to spare. With fail-safe gaps no crossing vehicle touched the failed one in 800 failures at the worst moment. Mitigated gaps let one arrive at up to 22 km/h, under the 25 km/h they allow, and nominal gaps at up to 35 km/h, close to the crossing speed. A vehicle that only brakes to a stop, the failure the nominal gap is sized for, was reached in 5 of 800 such failures with nominal gaps, at 12 km/h at most, and never with the wider ones.

  • Batches make most failures harmless to crossing traffic. At a random place and time, the next vehicle to cross the failed one's path is usually a whole phase away: its crossing is withdrawn, it stops short of the box, and it never needs the gap. Even with nominal gaps, 2 of 400 such failures led to a crossing contact, against 11% first come, first served, which needs its gap at almost every vehicle.

  • No gap protects the vehicle behind. In a lane, cars follow about 0.6 s apart at the crossing speed (a 2 m standstill gap plus a 0.42 s time gap), which lets them survive the one ahead braking, not stopping dead. In batches, a vehicle that stopped dead at a random moment was struck from behind in 86–89% of failures, at up to 38 km/h, whatever the gap. One that braked to a stop was struck 3 times in 4,800 failures, at 6 km/h at most, and never in batches.

  • The fail-safe gap has a blind spot. It is sized for the speed at which a vehicle enters a zone, and a vehicle still slowing from the approach speed to the crossing speed carries more than that when it has to react. First come, first served, an emergency stop ended up to 0.4 m inside its zone in 4 of 400 failures, clear of the failed vehicle. In batches, where the first vehicle of a phase isn't slowing as it enters, none did.

  • A disabled vehicle costs about a sixth of the flow while it stands, and the backlog takes minutes to clear. The manager closes the zones the vehicle holds and goes on serving the lanes that stay open. With the same arrivals and no failure for comparison, and the vehicle cleared away after 60 s:

    A vehicle stops dead in the box Served while it stands Most vehicles held up Back to normal after it is cleared away Delay added, per vehicle over 15 minutes
    Batches, 12,000 veh/h 84% 74 veh 3 minutes 2.4 s
    Batches, 16,000 veh/h 83% 106 veh 5 minutes 4.5 s
    Batches, 18,000 veh/h (design) 83% 121 veh over 13 minutes in 53 runs of 100 10.8 s
    First come, first served, 8,000 veh/h 88% 31 veh 1.5 minutes 1.0 s

To watch it, pick the cooperative crossing, turn on Compare side by side, give B batches and nominal gaps, and press Stall one at the worst moment.

Results

All tables: 20-minute runs after a 5-minute warm-up, realistic sensors (loops + camera + radar), default demand. Person delay* counts every vehicle that arrived, including those still queued, so a policy can't look good by starving an approach. Lower is better.

Controllers (bench all 20 6, 6 seeds)

Person delay* (s) Arterial Major/minor Town T-intersection
Fixed-time (Webster) 30.5 38.1 28.1 21.2
Actuated (NEMA) 37.5 57.5 26.5 18.3
Max-Pressure 20.4 23.2 30.2 15.8
Predictive MPC 16.9 25.4 21.8 14.8
Deep RL 21.3 26.0 29.0 12.3
Hybrid (RL-tuned predictive) 16.8 21.5 21.6 13.7
All-way stop 142.2 103.9 60.8 28.5
Two-way stop — 157.6 — 41.9
Roundabout 16.2

Zero vehicle overlaps in every run.

Training rewards every person's delay: people in vehicles on the road, in vehicles held at the edge of the map, and pedestrians waiting to cross. Episodes vary demand, weather and failed loop detectors. Checkpoints are chosen on 16 validation scenarios (design demand, 0.8× and 1.2× traffic, randomized demand, bad weather, a quarter of the loops failed), and a checkpoint is accepted only if its pedestrian waits stay within 5 s of the predictive controller's. The chosen weights are then scored on 24 test scenarios of the same kinds that no choice ever uses.

Both networks have one hidden layer of 128 units and were trained this way for 800 episodes. For Deep RL, one hidden layer beat two on every layout, by 1.8 to 4.5 s on held-out scenarios; for the hybrid, one and two did the same within noise. Held-out test set, person delay / pedestrian wait (s):

Predictive Best static tuning Hybrid Deep RL
Arterial 18.2 / 13.2 17.7 / 12.9 17.5 / 12.2 (network kept) 19.7 / 18.5
Major/minor 22.4 / 11.8 21.1 / 10.1 21.1 / 10.1 (static tuning) 23.0 / 16.3
Town 25.6 / 9.5 25.8 / 9.6 24.3 / 8.6 (network kept) 27.1 / 11.1
T-intersection 14.4 / 21.3 14.0 / 20.8 14.0 / 20.8 (static tuning) 12.9 / 19.9

Counting each waiting pedestrian like one vehicle occupant wasn't enough to protect pedestrians, who are a small share of the people at an intersection: Deep RL trained that way found a checkpoint that passed the pedestrian check in only 2 of 9 runs on the arterial, major/minor and T-intersection layouts. So Deep RL's safety shield now also gives the next walk to any pedestrian who has waited 40 s. With that rule every run passed, and the check never had to reject a checkpoint.

The rule costs people in vehicles little, and on most layouts it helps them. Deep RL trained three ways, held-out test set, mean of three training seeds, person delay / pedestrian wait (s):

0.3.0 New training, no rule New training + 40 s rule Predictive
Arterial 19.9 / 21.5 21.7 / 25.1 20.3 / 17.7 18.2 / 13.2
Major/minor 25.5 / 28.3 24.8 / 27.7 23.4 / 16.0 22.4 / 11.8
Town 27.3 / 14.8 28.7 / 11.8 26.7 / 10.7 25.6 / 9.6
T-intersection 12.0 / 45.6 11.5 / 36.8 12.6 / 19.7 14.4 / 21.3

With everything else the same, the rule cut person delay by 1.4 to 2.0 s on the four-way layouts as well as shortening pedestrian waits; only on the T-intersection did it cost 1.1 s. Against 0.3.0, person delay is within 0.6 s on three layouts and 2.1 s lower on major/minor, and pedestrian waits are 4 to 26 s shorter. A single network's benchmark score can mislead here: Deep RL networks trained from different seeds differ by 1 to 3 s.

Sensors (bench sensors all 20 4)

Arterial, person delay* (s):

loops + camera + radar + 30% CV 100% CV only ideal sensors
Max-Pressure 21.2 21.5 21.2 21.5 20.9 20.5
Predictive 20.5 17.5 17.3 17.5 17.3 18.0
Hybrid 20.9 19.0 17.3 18.1 17.9 18.9
Deep RL 30.3 21.0 21.8 21.7 21.6 21.5

Weather and failed loops (loops + camera + radar unless noted), predictive control:

Clear Night Rain Fog Snow loops only, 25% failed
Arterial 17.3 18.1 17.8 18.4 17.1 23.3
Major/minor 26.3 27.2 24.2 29.2 27.1 31.4
Town 23.1 29.5 26.3 22.8 29.6 45.3

Corridor (examples/corridor.rs, 8 seeds)

3 × arterial Person delay* Through-trip time Stops / trip
Fixed (isolated Webster) 30.8 199 s 2.24
Fixed (coordinated green wave) 26.7 166 s 1.92
Actuated 37.2 211 s 2.15
Max-Pressure 19.5 144 s 1.79
Predictive 17.8 137 s 1.51
Predictive + upstream link 17.9 140 s 1.55
Hybrid + upstream link 18.3 142 s 1.63
Deep RL + upstream link 19.6 142 s 1.52

On a 4-signal major/minor corridor the green wave cuts through-trip stops from 3.17 to 1.44 (time 249 s → 193 s). Predictive control is best overall (15.1 s); the upstream link hurts there (16.4 s). On the arterial it makes no difference at normal demand and little at 1.3× (36.4 → 35.5 s).

Signal-free cooperative crossing (examples/cooperative.rs, 3 seeds)

cooperative p2 2 mix 1200 3 14000,16000,17000,18000,19000,20000,22000: 20-minute runs after a 2-minute warm-up, no sensors, default vehicle mix, fail-safe gaps; each figure is the mean over the runs. "Served" counts vehicles that finished their trip; it trails the arrivals slightly while the approaches fill after the short warm-up.

Demand (veh/h) Arrived Served Average delay 95th percentile Longest wait Waiting to enter at the end
14,000 13,958 13,848 17.0 s 51 s 98 s 0
16,000 16,002 15,919 24.4 s 65 s 116 s 0
17,000 16,977 16,868 24.5 s 62 s 108 s 0
18,000 (design) 18,034 17,659 28.6 s 68 s 143 s 29
19,000 19,010 18,653 42.7 s 94 s 142 s 29
20,000 20,017 18,913 61.3 s 121 s 192 s 223
22,000 21,999 19,045 106.6 s 215 s 273 s 891

Capacity (served at 22,000 veh/h, or 24,000 for an all-car fleet) and the closest crossing measured. First come, first served serves the most at 14,000–16,000 veh/h of demand and slightly less as its queues grow (9,834 at 22,000), so its rows give that peak:

Default mix All cars Closest crossing (PET)
Batches, fail-safe 19,045 21,660 1.19 s
Batches, mitigated 19,386 0.68 s
Batches, nominal 19,702 22,329 0.50 s
Batches, fail-safe, one left-turn lane per approach about 13,500 1.13 s
First come, first served, fail-safe about 10,100 about 10,900 0.82 s
First come, first served, nominal about 12,600 0.50 s

Under first come, first served the closest crossings fall below a car's fail-safe gap at 40 km/h because the gap shrinks with speed, and vehicles that have slowed for their slot enter the zones slowly.

No run had a zone overlap, a box entry without a reservation or a step off a granted trajectory. Other arrival patterns at the design demand, 8 seeds each (cooperative p2 2 mix 1200 8 18000 platoons): platoons from an upstream signal average 34 s of delay, against 31 s for random arrivals on the same seeds. Random surges, which average 4% over the design demand with 2-minute bursts of up to 1.7× on an approach, average 48 s and leave about 160 vehicles waiting. The rush-hour peak (1.5×, about 27,000 veh/h) is far beyond capacity.

Failures at the cooperative crossing (examples/failures.rs, 400 runs per row)

failures 400: one vehicle fails in the box in each run, four minutes are simulated after it, and disabled vehicles are cleared away after 60 s. Batches at 16,000 veh/h, first come, first served at 8,000 veh/h. A failure comes at a random place in the box to the next vehicle that gets there, or at the worst moment: to a vehicle about to clear a zone with a crossing vehicle due right behind it at the smallest gap its policy allows. Contacts are the share of failures with one and the fastest closing speed. Stops and withdrawals are per failure. The closest stop is how far short of the zone it braked for the closest emergency stop of the 400 runs came to rest.

A vehicle stops dead Gaps Crossing contact Struck from behind Emergency stops Crossings withdrawn Closest stop
Random place and time, batches fail-safe none 89%, 38 km/h 2.5 22 1.3 m short
mitigated none 87%, 38 km/h 2.5 21 1.4 m short
nominal 0.5%, 29 km/h 86%, 38 km/h 2.5 21 6.9 m inside
Random place and time, first come, first served fail-safe none 26%, 38 km/h 1.0 11 0.3 m short
mitigated 3%, 14 km/h 15%, 38 km/h 1.1 9 4.1 m inside
nominal 11%, 34 km/h 8%, 38 km/h 1.3 8 9.9 m inside
Worst moment, batches fail-safe none none 5.6 39 0.6 m short
mitigated 85%, 21 km/h none 6.0 37 4.5 m inside
nominal 96%, 35 km/h 15%, 9 km/h 6.5 36 12.5 m inside
Worst moment, first come, first served fail-safe none none 2.0 15 0.4 m inside
mitigated 54%, 22 km/h none 2.2 12 6.6 m inside
nominal 89%, 35 km/h 3%, 18 km/h 2.7 13 12.8 m inside

A vehicle that brakes hard to a stop instead was never reached by crossing traffic with fail-safe or mitigated gaps, in 3,200 failures, and the closest emergency stop ended 0.8 m short of its zone. With nominal gaps it was reached in 5 of 1,600 failures, at up to 12 km/h. It was struck from behind 3 times in 4,800 failures, at up to 6 km/h, each time first come, first served.

The last emergency stop a failure caused was at rest 5.1 s after it at the latest: half a second of reaction time, then braking at 6 m/s² (4.5 m/s² for trucks and buses), and the same again for the vehicles that a later contact endangered. In all 9,600 runs no vehicle left the trajectory it held, entered the box without one, or overlapped another outside a recorded contact. Vehicles stopped in the box got in one another's way until one was towed 40 times, all first come, first served and all but one with mitigated or nominal gaps.

Pedestrian treatments (examples/lpi.rs, 12 seeds)

Near-misses = vehicles crossing within ~1.5 s of a walking pedestrian, per hour. At 240 pedestrians/h on the arterial: fixed-time 2.3 → 1.2 and predictive 2.3 → 0.8 with LPI (4 s) + no-turn-on-red-with-pedestrians, for 1–4 s more person delay. LPI alone helps little (1.8 and 2.5): most conflicts come from right turns on red and from turns later in the green. On the single-lane town layout neither treatment makes much difference.

Left-turn phasing (examples/fya.rs)

Protected + permissive (flashing yellow arrow) gives the least delay at normal volumes on the arterial; protected-only costs 20–76% more there. On the major/minor road fixed-time and actuated control still prefer permissive lefts, but predictive control does better protected-only (23.5 vs 29.4 s). At 1.3× demand on the major/minor road, fixed-time with permissive lefts collapses (157 s) as bays fill and spill into the through lane; protected-only holds at 98 s and Auto (volume cross-product) at 121 s.

What the experiments say

A fuller write-up of every finding, with the commands that reproduce it, is in docs/insights.md.

  • Model-based adaptive control is the strong default. Predictive MPC and Max-Pressure cut person delay 25–45% against an optimized Webster plan on the multi-lane layouts, and cut pedestrian waits by a fifth to a half on the four-way ones.
  • The best use of RL here is tuning a planner, not replacing it. The hybrid is the best or near-best controller on every layout, and it beats predictive control on every layout's held-out test set, by 0.4 to 1.3 s. Most of its gain comes from a better static tuning of the planner, which a one-off search finds. The state-dependent network adds 0.2 s on the arterial and 1.5 s on the town layout, and nothing on the other two. Pure Deep RL cuts person delay by about 30% against fixed-time on the multi-lane layouts and has the lowest delay on the T-intersection, but trails the best controller, the hybrid, by about 4.5 s on the multi-lane layouts and trails fixed-time and actuated control on the single-lane town layout.
  • Pedestrians need a hard rule, not just a share of the reward. Counting waiting pedestrians in Deep RL's reward left their waits long; a 40 s maximum wait in its shield fixed that on every layout, and on the three four-way layouts it lowered vehicle delay too.
  • Smaller networks learned better for Deep RL. One hidden layer of 128 units beat two on every layout and cut Deep RL's person delay by 10–12 s on the four-way layouts. For the hybrid, one and two layers did the same.
  • Coverage matters more than per-sensor accuracy. With realistic errors, a camera is still the most valuable single upgrade (intent and queues; town 47.7 → 22.5 s for predictive). Weather costs adaptive control 0–3 s on the multi-lane layouts because radar backs up the camera; night and snow cost about 6 s on the single-lane town layout. Failed loops matter a lot when loops are the only source. Diagnostics plus a volume-based fallback turn a collapse (~200 s) into graceful degradation (23–45 s).
  • Corridors: a coordinated green wave is worth having for fixed-time signals (stops per trip −14% to −55%). Adaptive control beats it without coordination. Naively sharing the upstream signal's outflow barely changes the arterial and hurts on the major/minor road; real gains need coordinated schedules, not just more lookahead.
  • Pedestrian safety: LPI plus a no-turn-on-red-with-pedestrians sign cut near-misses at busy arterial crosswalks by a half to two-thirds for a small delay cost.
  • Right tool for the job: with realistic reaction times an all-way stop is only competitive at low volumes (town at 0.6×: 15.0 s vs 10.5 s actuated), the roundabout is excellent at moderate volumes, and a two-way stop fails once the major road is busy.
  • Signal-free crossings still need phases. With every vehicle automated, an intersection manager carries the design demand of 18,000 veh/h at 29 s average delay and saturates near 19,000 veh/h with fail-safe gaps (19,700 with nominal gaps, about 21,700 if every vehicle were a car). The capacity comes from serving each movement in batches, like a signal without its yellow and all-red, not from weaving vehicles through one at a time: every switch between conflicting movements costs a clearance gap, and a weaving vehicle must find one in up to ten crossing lanes at once. Served first come, first served, the same crossing saturates near 10,100 veh/h; the zipper wins only below about 6,000 veh/h. The second left-turn lane on each approach is worth about 5,500 veh/h.
  • The safety gap is for the vehicle that crosses last. When a vehicle was made to stop dead in the box at the worst moment for crossing traffic, fail-safe gaps stopped that traffic short of it every time, mitigated gaps let it arrive at up to 22 km/h and nominal gaps at up to 35 km/h. In batches that moment comes only at a change of phase: at a random moment the next crossing vehicle is a whole phase away, and even nominal gaps almost never led to a crossing contact, while the zipper needs its gap at almost every vehicle. No gap protects the vehicle behind one that stops dead, which was struck in nine failures in ten mid-batch. A disabled vehicle costs the crossing about a sixth of its flow while it stands.

Repository layout

src/sim/        geometry, IDM + MOBIL vehicles, right of way, signals, sensors, demand, stats
src/sim/coop.rs   intersection manager for the signal-free cooperative crossing
src/sim/coop/failures.rs  a vehicle fails there: the reaction, contacts and recovery
src/sim/world.rs  worlds of one or more intersections (corridors, coordination)
src/control/    fixed, actuated, max-pressure, predictive MPC, deep-RL and hybrid controllers
src/rl/         MLP + Adam, Double DQN with replay
src/gui/        eframe/egui app: renderer, charts, panels, self-screenshot (TS_SHOT=…)
src/bin/        bench (comparisons, sensor study), train (RL training)
weights/        trained DQN and hybrid weights per layout
tests/          regression tests: safety, throughput, the app (headless) and the tools
examples/       corridor, lpi, fya, satflow, preempt, sweep (multi-seed safety sweep),
                cooperative, failures (a vehicle fails at the cooperative crossing)
docs/           architecture.md (how the code fits together, with a tour of the app),
                insights.md (every finding), cooperative-intersections.md (design notes for the
                signal-free crossing, with the scripts behind its estimates and figures),
                assurance-case.md (why SECURITY.md's expectations hold), screenshots

docs/architecture.md explains what each source file does, how one simulation step runs and how to add a controller, layout or sensor.

Notes and limitations

  • Corridors run along the east-west road only; side streets end at their own block.
  • The cooperative crossing is a single intersection. It assumes every vehicle is automated and connected, and it does not model human drivers, lost messages or a failed manager. A vehicle that is made to fail stops on its path, and the manager learns of it at once from the vehicle itself; one that couldn't report would have to be seen from the roadside, which isn't modeled. Vehicles that touch stop where they touch: nothing more of a collision is simulated.
  • The fail-safe gap is sized for the speed at which a vehicle enters a zone, so a vehicle still slowing to the crossing speed can end an emergency stop a few decimeters inside the zone (above).
  • The upstream link shares released vehicles only; it does not coordinate schedules between intersections the way SURTRAC does.
  • Drivers are well-behaved: no red-light running, distraction or crashes, so safety is measured with surrogates (overlaps, near-misses). The failures of the cooperative crossing are made on purpose.
  • Training is fast because the simulation runs thousands of times faster than real time per core: RESUME=1 cargo run --release --bin train all 300 900 hybrid fine-tunes the shipped weights.

ROADMAP.md says which of these the project plans to work on.

Three-signal corridor

Side-by-side comparison

Contributing

Bug reports, realism fixes and new experiments are welcome; see CONTRIBUTING.md and the Code of Conduct. ROADMAP.md says what's planned and what isn't. Every commit must be signed off under the Developer Certificate of Origin (git commit -s). GOVERNANCE.md describes how decisions are made and who does what. Please report security vulnerabilities privately, as described in SECURITY.md. ACCESSIBILITY.md describes what works with assistive technology, the known barriers, and how to report one.

License

Copyright (c) 2026 Jim Schreckengast.

Licensed under either of Apache License, Version 2.0 or MIT license at your option; LICENSES/ has the same texts under their SPDX names. Third-party components and their licenses are listed in THIRD_PARTY_LICENSES.md and ship with every binary package.

Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this project by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.

About

Microscopic traffic-intersection simulator in Rust. Compare fixed-time, actuated, max-pressure, predictive (MPC) and deep-RL signal control under realistic sensors, pedestrians, buses and corridors, and see the delay each person experiences. Runs in your browser via WebAssembly or on the desktop.

Topics

Resources

Code of conduct

Contributing

Security policy

Accessibility

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Used by

Contributors

Languages