Research / simulation ↔ reality
I study how well results transfer between simulation and the real world, how to make an autonomous-driving system more robust once that gap is measured, and how to reuse real driving data to make simulation more realistic.
Research themes
Measure which actuation, perception, and behavioral differences matter when a test moves between virtual and physical environments.
Found: Vehicle-in-the-loop testing reduced the actuation gap, and mixed-reality testing reduced the perception gap, compared with simulation alone.
Stress camera and LiDAR pipelines, then follow the effect beyond model accuracy into timing, planning, control, and system behavior.
Found: Offline robustness was only weakly predictive of closed-loop failures on the real vehicle; training on perturbed data improved robustness under adverse conditions.
Use near misses and predicted future states to spend the testing budget where failures are more likely to surface.
Found: Focusing the search around forecasted near misses found failures using a smaller testing budget than an unfocused search.
Turn recorded road scenes into replayable simulations that preserve the geometry, trajectory, and visual character of the original drive.
Found: It's not a full replacement for the real environment, but it lets us reuse real drives for closed-loop testing.
The interfaces and tooling that let the same test run in simulation, hardware-in-the-loop, or on a physical vehicle, without rewriting it each time.
Fortuna, the small-scale platform, and the tools around them are maintained so other students and researchers can keep using them, not just demo them once.
Seeing it work
Papers describe the method. These recordings show what it actually looks like.
Cam2Sim
Cam2Sim started from a practical problem. We had many hours of real driving recordings, but once an experiment was finished we could only replay the recorded sensor data. We wanted to turn these recordings into environments in which the driving system could interact again. Cam2Sim reconstructs the recorded environment in CARLA and uses Gaussian Splatting for the camera view. The reconstruction is not a complete replacement for the real environment, but it allows us to reuse real experiments for closed-loop simulation.
Read the paperASE 2025
An open-source small-scale vehicle and its Unity digital twin, compared against Fortuna at full scale, across simulation, vehicle-in-the-loop, and mixed-reality testing.
Read the paperPerturbationDrive
PerturbationDrive generates camera perturbations—like rain, or a corrupted feed—and applies them through a simulator adapter, so the same scenario runs unchanged across Udacity's simulator, a small-scale digital twin, or a full driving stack. This recording walks through the scenario generator and simulator adapter behind the ICST 2025 benchmark.
Read the paperForesee
Foresee monitors nearby agents, classifies and ranks the riskiest ones, then clips and fuzzes the scenario around them—so testing time goes where failures are most likely to surface.
Read the paperICST 2024
A recording of the ICST 2024 talk on assessing quality metrics for neural reality-gap input mitigation—testing whether standard image-quality metrics track how neural image translation affects a driving system's perception and control.
Read the paperBefore the reality gap
Before autonomous cars, I worked on autonomous nano-drones. At the Dalle Molle Institute for Artificial Intelligence (IDSIA) in Lugano, I helped fit a relative-localization neural network onto a 40-gram nano-quadrotor: a single onboard camera and under 100 mW of power, enough to track a peer drone in real time (IEEE ICRA 2023). The platform changed; the question didn’t: what has to be true for a learned model to work outside the lab it was trained in?
Read the paperA result is more valuable when someone else can interrogate it. The work is accompanied wherever possible by code, datasets, replication packages, and executable examples—from PerturbationDrive to Cam2Sim and vehicle-level artifacts.