NVIDIA AI Podcast Ep. 289

Title: Driving Safer AVs Faster with Smart Simulation, Neural Reconstruction, and Data-Centric Tools
Date: February 11, 2026
Host: Noah Kravitz
Guests:

• Rohan Basan, Senior Solutions Engineer, Fortellix
• Dan Garau, Head of Technical Partnerships & ML Evangelist, Voxel51

Episode Overview

This episode delves into how state-of-the-art AI-powered simulation, neural reconstruction, and data-centric tools are propelling the development of safer and more efficient autonomous vehicles (AVs). The conversation spotlights advances in simulation fidelity, the importance of "smart" data curation, the challenges of bridging physical and digital worlds, and how new tools are transforming developer workflows and team dynamics.

Key Discussion Points & Insights

1. Backgrounds and Roles

• Dan Garau (Voxel51): Provides advanced data platforms enabling AV teams to curate and target the most valuable data for model development—"garbage in, garbage out." Background in ‘physical AI,’ robotics, and edge devices.
  Quote: "If you’re not training on the right amount of data or the right data, you’re not going to be able to get that model that you’re hoping for." (01:19)

• Rohan Basan (Fortellix): Helps customers use sensor simulation tools for synthetic data generation, with expertise in simulation toolchains for driver assistance at Ford.

2. AV Simulation Fundamentals

• Shift From Modular to End-to-End Stacks:

  • Earlier focus was on separate neural nets for individual tasks (object detection, lane detection, etc.).
  • The field now moves toward integrated, generative AI-driven end-to-end stacks capable of outputting control signals directly (steering, throttle, braking).
  • Generative methods like 3D Gaussian splatting and diffusion models have drastically improved simulation fidelity. Quote (Rohan): "Now we see a drastic shift...to more end-to-end solutions...A big reason for that has been developments in generative AI." (02:51–03:47)

• Sensors in AVs:

  • Sensor stacks are highly varied (camera, LiDAR, radar, IMUs, physical sensors).
  • "It’s always a mixture...almost always camera data...then we call LiDAR data...then also very complex physics-based sensors." (Dan, 04:14–05:41)

3. Key Challenges: Real-to-Sim Translation & Data Gaps

• Quantity vs. Quality:

  • Having huge volumes of data (petabyte scale) isn’t sufficient; what matters is the right data and accurate translation from real-world to digital.
  • Dan: "If it was just more data, we would have self-driving cars." (06:15)

• Smart Data Use, Not More Data:

  • Fortellix helps devs identify what’s missing—targeting niche edge cases that are safety critical, rather than common, non-instructive scenarios.
  • Quote (Rohan): "Our goal is to help AV developers make sense of their data...fill in the gaps with exactly what they need, not scenarios of you pulling out of the garage but niche edge cases that are safety critical." (07:53)

4. Role of Foundation Models & Neural Reconstruction

• Neural Reconstruction:

  • Offers fidelity far superior to physics-based rendering.
  • Enables fast scenario variation (weather, anomalous pedestrian behaviors) without hand-crafting each case.
  • Reduces the need for real-world data collection or synthetic hacks (e.g., people crashing cars in GTA 5 for data five years ago!).
    Quote (Dan): "We literally had people playing GTA 5 and crashing into other people to capture data...now we can just imagine these things." (11:02–11:17)

• Time as the Crucial Resource:
  Quote (Dan): "The most valuable resource when you’re developing an AV system is not GPUs, it’s not data, it’s not people, it’s time." (09:41)

5. Testing, Edge Cases, and the 'Last Mile' Problem

• Edge Cases Remain Critical:

  • Unexpected situations (kids running into the street, sudden obstacles, rare weather) are the hardest to simulate but most important for safety.
  • Most nominal driving data is not incrementally useful.
  • "Anyone...can build a self-driving car that drives correctly 90% of the time. There’s just no way...you would ever release a self-driving car that only drives safely 90% of the time." (Dan, 14:01)

• Difficulty of Closing the Last Percent:
  Quote: "Each of those little jumps is like an order of magnitude more effort, more time, more data." (Rohan, 14:31)

6. Realism and Model Performance in Simulation

• Two Types of Realism:

  • For training: Does synthetic data improve end-to-end stack performance?
  • For testing: Does simulation replicate real-world failures?
  • "When you’re able to see new scenarios, add those...and hopefully learn not just that a person is running...but have the same reaction if it’s a dog or a cat..." (Rohan, 16:44)

• How Real Does Synthetic Data Need to Be?

  • Dan’s Devil’s Advocate: Absolute photorealism isn’t required if models learn effectively. After a certain level of realism, time is better spent generating more useful scenarios.
  • "The goal of training a self-driving car is to make sure it doesn’t crash...I actually don’t care if it doesn’t look exactly like the real world as long as my car gets better at driving." (Dan, 18:01)

7. Scenario-Driven Data Curation and Pipeline Automation

• Fortellix's Tools:

  • Automatically label temporal events in drive logs, letting teams search for and focus on overlapping or rare/critical scenarios (e.g., turning at intersections, yellow lights).
  • Scenario matching streamlines identifying and filling data gaps.
  • Quote: "Hundreds of petabytes of data...digging through that is a difficult process." (Rohan, 21:30–22:04)

• Voxel51's Data Curation:

  • Focuses on surfacing where a model feels "uncomfortable"—using embeddings and performance metrics to prioritize the most valuable scenes for reconstruction and further analysis.
  • "We need to hear and listen to our model to understand where we’re going to learn." (Dan, 23:08)

8. Collaboration, Visualization, & Workflow Enhancements

• Visualizations & Dashboards:

  • Essential for data trust, model performance analysis, and closing critical data gaps.
  • Tools let engineers quickly understand correlations between deficiencies (e.g., less data in parking lots => worse AV performance there).
    (Dan, 24:47–25:54)

• Continuous Feedback Loops:

  • Data translated, reconstructed, labeled, and fed back into the workflow—cyclical, continuously improving process.
    (Dan & Rohan, 25:55–27:09)

9. Interoperability & Data Language Challenges

• No Standard Common Data Language:

  • Each org has unique formats; forcing a single standard isn’t practical.
  • Key is accurate translation and alignment when passing data between modules/stacks ("not what language you’re speaking, but...that I’m translating it correctly"). (Dan, 27:53–28:34)

• World Models Change the Evaluation Pipeline:

  • End-to-end models allow for more holistic, less granular, evaluation—but root-cause failures become harder to identify.
    (Dan, 28:34–29:24)

10. Scaling on NVIDIA Platforms, Technical Advances & Business ROI

• Faster Rendering, More Variations:

  • Compute scale enables unprecedented scenario generation and faster development iterations.
  • Neural reconstruction leverages old data—logs that were previously useless now become valuable training/testing tools.
    "Thank goodness all this data we collected is not useless." (Dan, 30:23)

• Gaussian Splatting & Moving Objects:

  • Describes the leap from static 3D scene “painting” to moving-object-aware scenes—critical for realistic AV simulation. "That small advancement...has been probably the biggest step change that we’ve seen in the last six to eight months." (Dan, 32:23)

11. Voxel51’s Physical AI Data Engine

• Centralizes and standardizes data ingestion and curation for large, distributed, multi-team orgs.
• Involves rigorous auditing (alignment, sensor calibration, timestamps), automated enrichment (fixes, augmentation), and unlocks robust, cross-team reuse of key data improvements.
• "Probably more than 50% of the time, large companies that everyone would recognize, fail this test...data’s messy." (Dan, 35:14–36:44)

12. Future Outlook: Team Structure, Skills, and Workflows

• Flattened, More Cohesive Teams:

  • As automation increases, orgs may reduce specialized silos (simulation, reconstruction, safety, etc.), streamlining workflows and increasing collaboration.
  • "We can flatten these structures and roll them into one cohesive team...That sounds like it’s much more beneficial for everyone involved." (Dan, 39:54–40:14)

• Further Shift to World Models:

  • Coming era: models autonomously generate and test their own scenarios/variations, minimizing hand-crafted simulation.

• Faster Verification & Validation:

  • Neural reconstructions reduce reliance on real-world testing, enable faster feature rollout. Quote (Rohan): "It’s going to make trust in simulation a lot easier. That means V&V teams can sign off on new features...without as much extensive real-world testing and reducing costs." (38:56–39:37)

Notable Quotes & Memorable Moments

• On Data Quality Over Quantity:
  "If it was just more data, we would have self-driving cars." – Dan (06:15)

• On Time as the Limiting Factor:
  "The most valuable resource when you’re developing an AV System is not GPUs, it’s not data, it’s not people, it’s time." – Dan (09:41)

• On the Leap Forward:
  "We literally had people playing GTA 5 and crashing into other people to capture data...now we can just imagine these things." – Dan (11:02)

• On Photorealism vs. Effectiveness:
  "The goal of training a self-driving car is to make sure it doesn’t crash. I actually don’t care if it doesn’t look exactly like the real world as long as my car gets better at driving." – Dan (18:01)

• On the Last 1% of Safety:
  "Each of those little jumps is like an order of magnitude more effort, more time, more data." – Rohan (14:31)

• On New Workflows & Collaboration:
  "We can flatten these structures and roll them into one cohesive team...much more beneficial for everyone involved." – Dan (40:14)

Timestamps for Key Segments

• [01:07–02:18]: Guest introductions and company missions
• [02:51–03:47]: Evolution of AV simulation and end-to-end solutions
• [04:14–05:41]: Sensor types and real data sources in AVs
• [06:15–07:53]: Data challenges from real-world to digital translation
• [08:54–11:17]: Impact of neural reconstruction and saving development time
• [12:40–14:01]: The challenge of AV edge cases and closing the safety gap
• [15:06–19:22]: Measuring realism in simulation; photorealism vs. effectiveness
• [20:38–22:04]: Fortellix’s scenario-driven curation and Fortify suite
• [22:15–24:30]: Voxel51’s data curation and reconstruction pipeline
• [24:47–27:09]: Visualization, feedback loops, and collaborative cycle
• [27:40–29:24]: Challenges with data interoperability and world model evaluation
• [29:48–32:38]: Technical advances with NVIDIA platforms: scale, rendering, Gaussian splatting
• [33:05–38:02]: The Physical AI Data Engine, cross-team data workflow, and impact
• [38:36–42:01]: Future of team structures, skill needs, and world model-driven simulation

Further Resources

• Voxel51 Docs & Tutorials
• Fortellix Official Site
• Contact Dan Garau on LinkedIn
• Contact Rohan Basan on LinkedIn

Final Note

The episode provides a dynamic overview of how AV teams are leveraging the latest AI tools to simulate, test, and improve safe driving performance—emphasizing time savings, smarter data selection, and the shift to autonomous, data-driven modeling—all at an unprecedented pace.