In this livestream, we will present a general, theoretically grounded framework for AV safety validation whereby real-world tests are paired with simulated tests on corresponding reconstructed scenarios. We will discuss how this approach can dramatically reduce the number of real-world tests required to estimate safety KPIs with a specified level of confidence.
Attendees will learn:
· Main considerations to validate high-end autonomy stacks.
· How simulation can be used to significantly accelerate the task of AV safety validation.
· More broadly, opportunities offered by generative AI to instantiate a safety data flywheel.
Key Moments:
00:19 — Welcome & agenda: AV safety validation via simulation
04:03 — Why end-to-end AV stacks need new safety validation (foundation models)
06:37 — Three pillars of safety: Diversity, Monitoring, Evidence
15:12 — NVIDIA approach: end-to-end policy + modular guardrails for fail-safe behavior
18:21 — Car2Sim: neural scene reconstruction (Cosmos) to re-simulate real logs
20:28 — SIM-to-VAL method: control-variates pairing of real + sim with sim-only boosts
32:39 — Case study (intersections): ~6× fewer real-world miles for same confidence
40:41 — Demo results: 34.5% variance reduction vs. vanilla Monte Carlo
42:57 — Metric Correlator Function (MCF): learning better sim→real correlation
44:06 — Key takeaways: sim as variance-reduction, not sim-only validation