JunctionTwin · Traffic simulation

Stress-test a junction before you build it.

A road junction is poured once and argued about for a decade. JunctionTwin rebuilds it digitally first, fills it with the traffic that actually uses it, and pushes the design until it fails — so the failure happens on a workstation instead of on the road.

01

Why this problem

Our roads break the assumptions most simulators make.

Established traffic tools model lane-following vehicles on well-regulated networks, and they do it well. The two groups most at risk on Indian roads are precisely the ones those models represent worst.

480,583

Road accidents recorded across India in 2023 — up 4.2% year on year.

172,890

Lives lost in a single year, 83.4% of them working-age adults.

44.8% + 20%

Two-wheeler riders and pedestrians as a share of those deaths.

Source: Ministry of Road Transport & Highways, Road Accidents in India 2023.

02

What it does

Six steps, from drawing to evidence.

Rebuild

Reconstruct a real junction — geometry, lanes, approaches, signal logic — as an OpenUSD scene.

Populate

Fill it with agents calibrated to observed behaviour: two-wheelers, autos, trucks, pedestrians, livestock.

Break it

Run peak demand, incidents, rule-breaking and edge cases until throughput or safety collapses.

Measure

Flow, queue length, delay, conflict points, time-to-collision and lane utilisation — numbers, not impressions.

Compare

Test alternative layouts side by side and quantify the difference instead of arguing about it.

Report

Deliver an evidence pack an engineering or municipal team can act on the next morning.

The hard part

The question is not whether the design looks good.

It is whether it survives a Tuesday at 6pm.

A simulation that only models well-behaved vehicles will approve a junction that fails in its first week. Every behaviour listed here is routine on the roads we intend to model — and an edge case for a simulator built elsewhere.

  • Pedestrians crossing mid-block, against signals, in groups
  • Animals entering or standing in the roadway
  • Two-wheelers filtering between lanes and through gaps
  • Drivers ignoring lane markings and right-of-way
  • Late merges, hesitation and sudden stops
  • Breakdowns, collisions and blocked movements
03

How a design earns approval

Six regimes, each producing numbers rather than opinions.

A design that fails any of them goes back to the drawing board — on a workstation, not on a construction site.

  1. Regime 01

    Baseline demand

    Does the design improve flow when traffic is comfortable?

  2. Regime 02

    Peak demand

    Where do queues form first, and how fast do they propagate upstream?

  3. Regime 03

    Demand +25–50%

    At what growth level does the design lose capacity entirely?

  4. Regime 04

    Human behaviour

    What happens under hesitation, late merging and poor lane discipline?

  5. Regime 05

    Heavy vehicles

    Do trucks and buses create blind-spot conflicts with two-wheelers?

  6. Regime 06

    Incidents

    What does a single blocked lane do to the whole junction?

04

How it is built

Accelerated computing, because the sweep is the product.

One simulated afternoon proves nothing. The value is in running hundreds of variants — layouts, demand profiles, driver behaviours, incidents — and that only becomes affordable on GPU.

OpenUSD scene graph

Junction geometry as a physically accurate, interoperable scene — the digital twin itself.

Synthetic data

Domain randomisation over lighting, weather, camera placement and vehicle mix, so models see more than one afternoon's footage.

Generated edge cases

Rare and dangerous traffic situations that are unsafe — or impossible — to capture from real recordings.

Video → trajectories

Junction camera feeds turned into vehicle counts, classes and paths, which is what the model is calibrated against.

Behaviour models

Detection and trajectory models fine-tuned on synthetic plus local footage, then evaluated against held-out days.

Parallel sweeps

Hundreds of design variants and demand profiles evaluated per run on GPU, rather than one scenario at a time.

05

Where we actually are

Stated plainly, because you will find out anyway.

Built

Three production AI products serving real users, a decade of Unity and real-time 3D work, and in-house computer-vision and LLM engineering.

Not built

JunctionTwin itself. No simulation code, no calibrated junction, no pilot. This is a concept backed by a team, and we are not going to dress it up as more.

Next

One engineer is working through the simulation and synthetic-data stack. Junction selection and traffic-data sourcing in Hyderabad are the immediate step.

06

The first 90 days

A plan with a milestone that can fail.

  1. Days 1–30

    Pick one junction

    Select a real junction, collect geometry, traffic counts and 20+ hours of video, and build the scene.

  2. Days 31–60

    Calibrate against reality

    Detection and tracking on the footage, then tune agent behaviour until simulated flow matches measured flow.

  3. Days 61–90

    Try to break it

    Run the six-regime stress matrix across three layout variants, and publish the results — including the bad ones.

The test we have set ourselves

Simulated baseline flow within ±10% of measured flow at the real junction, across three separate time windows.

If calibration cannot get inside ±15%, we will publish that too and revise the approach before going further. A model nobody has tried to falsify is a rendering, not a simulation.

Let's build

Have something worth building?

Tell us who it helps. If it makes people better off, we'd love to build it with you.

Start a project