End-to-End LiDAR optimization for 3D point cloud registration

LiDAR sensors are a key modality for 3D perception, yet they are typically designed independently of downstream tasks such as point cloud registration. Conventional registration operates on pre-acquired datasets with fixed LiDAR configurations, leading to suboptimal data collection and significant computational overhead for sampling, noise filtering, and parameter tuning. In this work, we propose an adaptive LiDAR sensing framework that dynamically adjusts sensor parameters, jointly optimizing LiDAR acquisition and registration hyperparameters. By integrating registration feedback into the sensing loop, our approach optimally balances point density, noise, and sparsity, improving registration accuracy and efficiency. Evaluations in the CARLA simulation demonstrate that our method outperforms fixed-parameter baselines while retaining generalization abilities, highlighting the potential of adaptive LiDAR for autonomous perception and robotic applications.

💡 Research Summary

The paper introduces an end‑to‑end adaptive framework that jointly optimizes LiDAR acquisition parameters and point‑cloud registration hyper‑parameters, addressing the long‑standing separation between sensor design and downstream perception tasks. The authors argue that conventional pipelines treat LiDAR as a static data source, which leads to sub‑optimal point density, noise levels, and ultimately degraded registration performance. To close this gap, they propose a closed‑loop system where registration quality directly feeds back to adjust LiDAR settings such as pulse power, gain, beam divergence, saturation level, and pulse width.

The technical core of the method is a unified optimization problem that minimizes the registration error (both translation and rotation) over a set of pose pairs. The LiDAR forward model f_lidar takes a 6‑DoF pose and a vector of sensor parameters Θ_lidar and produces a point cloud by simulating full transient waveforms, applying analog‑to‑digital conversion, peak detection, and clipping. The registration function f_reg consumes two point clouds generated with the same Θ_lidar and a set of registration parameters Θ_reg (voxel grid size, FPFH normal and feature radii, nearest‑neighbor count, correspondence distance threshold, and ICP iteration count).

Because the search space includes both continuous and discrete variables with highly non‑linear interactions, the authors employ Covariance Matrix Adaptation Evolution Strategy (CMA‑ES). This evolutionary optimizer adapts a covariance matrix that captures parameter correlations, allowing it to discover, for example, that increasing LiDAR power (which raises point density but also noise) should be compensated by enlarging FPFH radii to preserve descriptor distinctiveness. All parameters are normalized to