Sim-to-real validation for robot fleets

Prove every robot update before it ships

Simprova runs each new policy through thousands of dangerous scenarios inside a digital twin of your own site. You get a signed pass or fail before a single robot downloads it.

Built on NVIDIA Omniverse, Isaac Sim and Cosmos. Cloud or on-prem.

Example certificatePass

pick-policy v4.2.1

Site
Warehouse B twin
Scenarios run
12,480
Collisions
0
Min. human clearance
0.62 m
p99 stop time
212 ms
Est. incident rate (95% CI)
< 1.3 × 10⁻⁶ / h
Human intrusion100%
Sensor dropout100%
Floor and surface99.6%
Occlusion100%
Hardware faults99.8%
Illustrative values. Real certificates are signed, versioned and tied to one policy build and one twin version.

What every update gets

One gate between your model and your fleet

Twin1

Calibrated digital twin per site, rebuilt from your own scans and telemetry.

Parallel rollouts1,000s

Randomised runs per update on data-centre GPUs, finished in hours.

Certificate1

Signed pass/fail with a quantified risk estimate and full evidence.

Audit trail100%

Every run, seed and outcome logged and replayable.

The problem

Robots now update like software. They can't be tested like software.

Learned policies ship over the air every few weeks. A bad update can damage stock, stop a line or hurt a person, and testing on physical robots covers a tiny fraction of the cases that matter.

Most teams end up choosing between shipping slowly and shipping blind. Simprova removes that trade-off.

  • Physical test cells see hundreds of cases, not thousands
  • Generic simulators miss your floor, racks, lighting and traffic
  • There is no evidence trail when a safety lead or insurer asks

Without a gate

Update → fleet → incident report

With Simprova

Update → twin → 12,480 runs → verdict → fleet

The platform

Five products, one closed loop

Each stage feeds the next, and real-world telemetry flows back so the twin and the scenario library get closer to reality with every deployment.

LoopSentryx telemetry recalibrates Twinora and adds new cases to Scenova, so every update is tested against more of reality than the last.

How a release gets proven

From pull request to rollout in four steps

  1. Minutes

    Submit the build

    Your CI pipeline sends the candidate policy to Simprova with one CLI call or webhook.

  2. Hours

    Sweep the twin

    Simvolt runs it across the site twin and the scenario suite you chose, in parallel.

  3. Instant

    Read the verdict

    Veridex scores every run and issues a signed pass or fail with the failing cases attached.

  4. Your call

    Roll out or fix

    A pass unblocks your OTA rollout. A fail ships replays your team can debug.

Built on the NVIDIA stack

GPU simulation where it matters

Simprova is GPU-native from the twin renderer to the edge agent.

Omniverse + OpenUSDDigital twins
Isaac Sim / Isaac LabParallel robot simulation
CosmosWorld-model scenario variation
DGX / HGXSweep compute
TensorRTFast policy inference
Jetson Orin / IGXEdge telemetry agent
HoloscanSensor streaming
NVIDIA AI EnterpriseSupported deployment

See every integration →

Developer first

A release gate that lives in your pipeline

Simprova plugs into the CI you already run. Fail the build when the verdict isn't a pass, and attach the certificate to every release.

  • CLI, REST API and webhooks
  • GitHub Actions, GitLab CI and Jenkins examples
  • ONNX, TorchScript and TensorRT policy formats
release.yml
# .github/workflows/release.yml — gate a policy release on a Simprova certificate
jobs:
  validate:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - name: Submit policy for validation
        run: |
          simprova sweep create \
            --policy ./build/pick-policy.onnx \
            --twin site-warehouse-b@latest \
            --suite release-gate \
            --wait --fail-on verdict!=pass
      - name: Attach certificate to the release
        run: simprova cert download --sweep $SWEEP_ID -o cert.json

Questions

Frequently asked

Do we need to change our robot software?

No. Simprova tests the policy artifact you already build. The edge agent is optional and runs alongside your stack.

How long does it take to build a twin of our site?

A pilot twin is planned in two to three weeks from CAD, LiDAR or photogrammetry scans plus a sample of fleet logs. It then keeps calibrating itself from telemetry.

Is a Simprova certificate a regulatory certification?

No. It is engineering evidence that supports your own safety case and audits. It does not replace a notified body or your risk assessment.

Can it run on-prem?

Yes. Simprova runs in our cloud, in your VPC, or on your own DGX hardware for sites that can't send data out.

See your own site in a twin

Bring a floor plan and a recent policy build. We'll show you what a sweep finds.