NVIDIA AI Podcast Driving Safer AVs Faster with Smart Simulation, Neural Reconstruction, and Data-Centric Tools - Ep. 289
65 snips
Feb 11, 2026 Dan Gural, head of technical partnerships at Voxel51 focused on data curation and workflows. Rohan Bhasin, senior solutions engineer at Fortellix specializing in sensor simulation and synthetic data. They discuss turning drive logs into high-fidelity simulations, neural reconstruction for faster scenario generation, and data-centric tools that prioritize edge cases and scale AV validation.
AI Snips
Chapters
Transcript
Episode notes
Shift To End-To-End Simulation
- The AV simulation ecosystem shifted from camera-only tests to end-to-end stacks combining perception and control.
- Advances like generative AI and neural reconstruction raised expectations for richer sensor simulation data.
Quality Over Quantity Of Data
- More data alone doesn't solve AV safety; translation quality from physical logs to digital matters most.
- Fortelix and Voxel51 focus on accurate translation and targeted scenario generation to close performance gaps.
Generate Edge Cases From Existing Logs
- Use neural reconstruction and foundation models to create realistic variations (weather, actor behaviors) from existing logs.
- Generate edge-case situations instead of collecting vast new datasets to save time and improve training.