Luca Brembilla

I came to autonomous-driving perception through my MSc thesis at Politecnico di Milano. I was trying to make LiDAR detection fast enough on an edge device, and developed an ROI strategy that roughly halved the points processed. The work became an ICIAP 2025 paper.

I then joined AIDA’s perception team, first as a research engineer and now as a PhD researcher advised by Matteo Corno. I train and benchmark 3D detectors, including camera–LiDAR models, and work on getting them into our vehicle stack within its hardware and latency constraints. The perception software I wrote runs across the AIDA/NiuLinx vehicle fleet. NiuLinx is a €38M-funded spin-off developing autonomous-driving technology in Italy and Europe. I am also an authorized supervisor for autonomous-driving tests on public roads.

My current research is about self-supervised pre-training for perception. In What Does the Future Buy?, I separate the effects of feature transformation and future supervision in LiDAR pre-training. In the detection settings I tested, the gains come from the transformation rather than a stable benefit from future targets.

I am also building DriveRT, an open-source compiler and standalone runtime for sparse 3D detectors. It compiles a supported PyTorch model into a single file that an existing application can run through a small C library, without Python or PyTorch at runtime. VoxelNeXt is the first supported detector; I am extending it to more models and precision modes.