Abstract
<title>Abstract</title> <p> Engineers and researchers suffer from current LiDAR perception: each new sensor typically requires a dedicated dataset, model, and training pipeline, and each team repeats the costly process, as data are usually not shared, severely hindering deployment and increasing R&D costs. This limitation largely stems from the prevailing single-dataset training paradigm, under which models easily overfit to sensor-specific sampling patterns and generalize poorly to unseen LiDARs. This work tackles this problem from a new perspective: <italic>any LiDAR can be viewed as a sparse sampler of the underlying 3D world</italic> , so each scan can be decomposed into scene-intrinsic geometry (what the scene is) and sensor-specific sampling effects (how the sensor samples it). Based on this insight, we propose LiSeer, a unified pipeline ''learning from randomness'' for zero-shot cross-LiDAR generalization. LiSeer first reconstructs the underlying scene using a realistic diffusion-based densification model, and then applies a physics-based controllable sampling strategy to mimic diverse LiDAR point clouds. This enables virtually unlimited generation of cross-sensor data and alleviates the data-hungry bottleneck. Building on this foundation, we further encourage the model to prioritize scene-intrinsic geometry over sensor-specific sampling effects through explicit scene-sensor disentanglement and consistency learning. Extensive experiments across three datasets and two mainstream segmentation backbones show that LiSeer delivers large, consistent zero-shot gains, improving baseline cross-sensor mIoU by 25.1 to 43.6 percentage points. In our experiments, scaling the generated data further improves accuracy. With only 10%-20% of the target-domain data, it approaches fully-supervised models. Deployed on real vehicles with previously unseen 32-, 80-, and 128-beam LiDARs, LiSeer produces accurate and stable zero-shot segmentation, providing strong evidence of its practical value. </p>