← Back ICRA 2026

LiLa-Net: Lightweight Latent LiDAR Autoencoder for 3D Point Cloud Reconstruction

Mario Resino Solis, Borja Pérez López, Jaime Godoy Calvo, Abdulla Al-Kaff, Fernando Garcia

PDF

AI summary

Key figure (auto-extracted from paper)

LiLa-Net enables efficient, high-fidelity 3D LiDAR point cloud reconstruction and cross-dataset generalization using a lightweight autoencoder with optimized skip connections.

LiDAR Point Cloud Autoencoder Lightweight 3D Reconstruction Autonomous Driving

Problem

Real-time autonomous driving requires efficient processing of high-volume LiDAR point clouds, but current Transformer-based and masked autoencoder methods are too computationally expensive for practical deployment.

Approach

The authors propose LiLa-Net, a lightweight end-to-end autoencoder that directly processes sparse 3D points using a reduced encoder depth and a single optimized skip connection to balance latent encoding and reconstruction fidelity.

Key results

Direct processing of sparse 3D points without voxelization or masking
Accurate reconstruction of complex traffic environments with minimal error
Strong generalization to unrelated 3D object datasets without retraining
Faster inference and smaller model size compared to state-of-the-art architectures

Why it matters

Offers a practical, resource-efficient solution for real-time 3D perception and feature extraction in autonomous vehicles and robotics.

Abstract

This work proposed a 3D autoencoder architec- ture, named LiLa-Net, which encodes efficient features from real traffic environments, employing only the LiDAR’s point clouds. For this purpose, we have real semi-autonomous ve- hicle, equipped with Velodyne LiDAR. The system leverage skip connections concept to improve the performance without using extensive resources as the state-of-the-art architectures. Key changes include reducing the number of encoder layers and simplifying the skip connections, while still producing an efficient and representative latent space which allows to accurately reconstruct the original point cloud. Furthermore, an effective balance has been achieved between the information carried by the skip connections and the latent encoding, lead- ing to improved reconstruction quality without compromising performance. Finally, the model successfully reconstruct objects unrelated to the original traffic environment.

Index terms

Deep Learning Methods Intelligent Transportation Systems