Pre-Silicon System Verification

WorldSim: Industrial Closed-Loop Virtual Proving Ground for AI Silicon

WorldSim runs AI silicon, firmware and model weights inside a virtual industrial campus before tape-out. Roads, a factory gate, a yard, a warehouse and a loading dock share one simulation clock, and every run leaves a record that can be replayed.

Factory scenario, one recorded run, 20 seconds. A child runs out from behind a pillar. The firmware lane panel shows a perception firmware image running on the SoC virtual platform. The decision ledger shows the firmware decision and the action it produced. The referee panel holds ground truth, which the driver does not receive. The inset in the lower right is a rear camera view.

The reason

Why verify hardware in a virtual world

Silicon faults often appear only when hardware reacts to a changing scene: a decision that arrives a few frames late, a sensor frame that is dropped, a firmware branch that no bench test reaches. WorldSim applies one scenario to the sensor model, the virtual silicon and the vehicle model in the same loop, so a late decision is visible in the same run as the stop it causes.

Each run is defined before it starts

A run is set by a seed, a scenario variant, and the hashes of the build. Simulation state advances on integer ticks, and a lint check blocks wall-clock reads in that code.

A result can be checked after the fact

The run writes fingerprints and a hash-chained journal. When two runs disagree, the first tick where they differ can be found, and the journal entry at that tick shows what changed.

One continuous industrial world

One truck, one clock, one registry

The campus is a synthetic 3 km² site with 23 junctions, 76 crossings and 6 warehouses. The environment is synthetic, designed from industrial campus patterns rather than reproduced from a specific facility. Everything on it runs on one simulation clock with one registry of actors, so a truck keeps its identity as it moves from the road into the yard, the warehouse and the dock.

  1. Road approach: roadmap Traffic: measured

    Public road and signalised junctions

    Vehicles and pedestrians run on the campus network under signal, give-way and crossing rules. The highest tested passing rate is 120 vehicles per hour per gate under the stated demand mix and horizon. That is a result of this test, not a figure for a real facility.

  2. Gate: measured

    Factory gate

    Two gates, 18 dock controllers and 17 signal controllers run as entities in the same registry. A dock release is refused unless every required proof is present. Six injected controller faults each reached a safe state.

  3. Yard: measured

    Yard and reverse approach

    The truck follows a 702 m road, enters the yard and performs a 40 m closed-loop reverse toward its bay.

  4. Warehouse: measured

    Warehouse and forklifts

    Six warehouses run pallet jobs with forklift fleets. Forklifts reserve the aisle lanes they enter. Fleets of one to four forklifts completed 18, 30, 36 and 42 job cycles, with no stall.

  5. Dock: measured

    Loading dock

    The dock bay runs through its states from assignment to completion. The interlock chain is checked after every transition, and the truck is released only after the trailer is clear of forklifts.

The measured mission starts at the factory gate. The public road approach is listed on the roadmap and is not part of the measured run yet.

Five claims

The five claims and their evidence

Each claim states what WorldSim shows today, the record that shows it, and its current limitation. Records are named by date and check so they can be requested for review.

Claim 1

World continuity

One truck, one clock and one registry carry a mission from the gate through the yard and the dock to departure.

Shown today
A one-hour simulated shift of the truck mission ran 7 cycles and moved 32 pallets out. Every pallet is named in the census, none are unaccounted for, mass is conserved, and no deadlock was detected. Four runs with the same seed gave the same fingerprint, and a replay matched byte for byte.
Evidence
Campus mission record, 2026-09-11, item 6. The mission gate passed 52 of 52 checks.
Current limitation
The measured mission starts at the gate, not on the public road. Sampled renders use simple box geometry, not finished assets.

Claim 2

Sensor generation

Camera frames and their labels come from the same world state, and the driver never receives ground truth.

Shown today
Each frame carries instance IDs, masks, depth at 1/64 m resolution, visibility and track IDs. Labels match the world's ground truth to a residual of 2e-4 px, with zero track ID swaps in the test. Ground-truth frames and sensor frames are separate types, and a runtime guard refuses any outgoing frame that carries a ground-truth field. In a scramble test, a clean planner gave bit-identical results with ground truth scrambled.
Evidence
Label audit record, 2026-09-12. Sensor layer checks: 25 for configuration and 31 for the ground-truth firewall.
Current limitation
Sensor noise models are specified but not built. Camera configurations are not yet calibrated. Transport separation of sensor and ground-truth streams is not enforced at deployment. Campus-render perception accuracy is not yet claimed. Visualization overlays remain engineering instrumentation, not production UI.

Claim 3

Virtual-silicon execution

Firmware and model weights run on virtual silicon inside the loop, and each decision enters the simulation at a recorded tick.

Shown today
Perception firmware images run on a SoC virtual platform. The firmware lane is asked for a decision every 60 ticks (2 Hz), and each decision is journaled with the firmware identity and a stdout hash. Neural network models run in the same loop, with their weight hashes recorded at load. The RTL lane has passed 450 of 450 testbenches on its vector-in path.
Evidence
Inference matrix record, 2026-09-12: 208 of 247 cells produced, each with model and preprocessing identity bound. RTL testbench record for the vector-in lane.
Current limitation
Firmware operating point remains under evaluation. No camera frame reaches an RTL decision yet, so RTL is not in the closed loop.

Claim 4

Closed-loop control

Each decision changes the vehicle, the vehicle changes the next sensor frame, and the referee judges the result.

Shown today
Four legacy factory exams re-hosted on the campus gave the same verdicts as the factory scene in 3 of 3 comparisons, with stop distances inside tolerance. Replay from the journal reproduced the fingerprint in 4 of 4 cases. A decision that arrives outside its freshness window is refused: in one test, stale frames were offered 104 times and acted on 0 times. In the forced case with a forklift in the trailer, the dock interlock refused release.
Evidence
Campus exam record, 2026-09-11, item 4. Decision freshness record, 2026-09-11. Mission record, 2026-09-11, item 6.
Current limitation
The actuated signal controller cannot see an exam ego, so the graded exams declare fixed signals. No closed-loop pass rate is claimed on held-out families yet. The driving gate does not yet separate the five detector lanes it was run on.

Claim 5

Deterministic replay evidence

The same seed, scenario and build give the same run, and a difference in one tick can be located by bisection.

Shown today
A 144,000 tick run, which is 20 minutes of simulated time at 120 Hz, with 500 entities keeps 120 keyframes and a hash-chained journal in 6.13 MB. When one pose was changed by 1e-9 at one tick, bisection found that tick in 10 comparisons. Injected 500 ms wall-clock stalls left the final fingerprint unchanged. On the campus, state determinism measured 0.000000000 mm. Nine math functions returned bit-identical results over 10,000 inputs each across several runtime engines and an independent reference implementation.
Evidence
Simulation core record, 2026-09-08 (stall, registry and keyframe tests). Campus integration record, 2026-09-11 (math vectors and state determinism).
Current limitation
Replay shows that the same inputs give the same run. It does not show that the model matches a physical vehicle. Clock alignment between GNSS, inference timestamps and recorded video is not yet validated across runs.

How a run works

From scenario to evidence

Each step below is the stage that the claims above depend on. The order is the order in which a single run executes.

  1. World

    A seed selects the scenario variant. The campus, the road segment and the warehouses load from data files at the start of the run.

  2. Sensors

    Camera frames and ground-truth frames are rendered from the same tick. Only sensor frames leave the simulator toward the brain.

  3. Virtual silicon

    Firmware or a model receives sensor frames at its declared cadence and returns a decision. The firmware identity and output hash are journaled with it.

  4. Decision

    The decision is applied at a recorded tick. A decision that arrives outside its freshness window is refused and counted.

  5. Referee

    The referee checks the scenario rules against ground truth. The driver never receives that data.

  6. Evidence

    The run writes fingerprints, the journal and a run bundle. Replay and bisection read the same records, so a verdict can be checked again.

Campus mission clip

A truck mission on the campus

The clip shows the campus view from a follow camera while the truck reverses toward a dock bay. The status overlay lists the current stage, the bay and dock state, the tick, and the counts of junctions, driving lanes, buildings, docks, forklifts, pedestrians and vehicles in the run.

Campus mission, 30 seconds. A truck reversing toward a warehouse dock, seen from a follow camera.

Roadmap

Not built yet

Roadmap: not working today
  • A measured mission that starts on the public road and passes the signalised junction before the factory gate.
  • Camera noise models, and calibration for each camera rig.
  • A camera-frame-to-decision path through the RTL lane, so that RTL decisions run inside the closed loop.
  • Perception accuracy on campus renders, which needs photoreal assets first.
  • A firmware operating point, chosen on held-out scenario families.
  • The ground-truth scramble test in continuous integration, and sensor and ground-truth transport separation enforced at deployment.
  • A recorder clock audit that aligns GNSS, inference and video timestamps and checks drift across runs.
  • Seed variation that changes ambient traffic. Current seed sweeps do not vary it.
  • Plant calibration against measured braking and steering, and a road exam with a referee scorecard on recorded drives.

Evaluate WorldSim

Bring the verification question

Tell us the design stage and the verification question you want WorldSim to answer. The enquiry form asks for your name, company email, organization, role, country, enquiry type, project stage, timeframe, and a short description of the requirement. WIOWIZ logs each submission, and the inquiries team reviews it and replies by email.