Luca Brembilla

I started working on autonomous-driving perception during my MSc. I was trying to make a LiDAR detector run fast enough on an edge device, and ended up developing an adaptive crop that removed half the points and doubled the frame rate. That work became my thesis and an ICIAP 2025 paper.

After that, I joined AIDA’s perception team. Most of my work sat between the model and the vehicle: training 3D detectors, integrating them into ROS 2, and fixing the TensorRT and CUDA pieces needed to run them in real time. A CUDA implementation of PointCloud2 unpacking, for example, brought that step from 9 ms to 1.5 ms. Some of the perception software I wrote at Politecnico di Milano was later transferred to NiuLinx.

Now I am doing a PhD with Matteo Corno and Giacomo Boracchi. I study self-supervised learning for LiDAR and camera–LiDAR models. More specifically, I want to know what pre-training adds to their representations, and how much of it is still useful after fine-tuning. Outside that work, I am building DriveRT, an open-source compiler and runtime for sparse 3D detectors.