Original title: Toward Robust LiDAR Semantic Segmentation for Real-World Deployment: Evaluation under Coarse Labels, Adverse Conditions, and Domain Shifts
Samir Abou Haidar, Alexandre Chariot, Mehdi Darouich, Cyril Joly +1 · September 2026
Self-driving cars and mobile robots use laser sensors called LiDAR to see and label their surroundings. Today's top models score high on standard benchmarks, but these tests use clean, single-domain data and overly detailed labels. In reality, cars face messy weather, hardware glitches, and brand-new locations. The researchers built a structured evaluation protocol to test how ready these systems truly are. They tested models across three dimensions: coarse labels that matter for driving safety, eight types of real-world sensor corruptionsArtificial modifications to data that simulate real-world issues like weather or sensor errors. like atmospheric and geometric noise, and domain shifts across datasets without adaptation. They also clocked processing speeds on an embedded Jetson AGX OrinA specialized computer chip used for running AI in embedded systems like self-driving cars. platform. The results reveal that high benchmark ranks do not mean a model is safe. Every method degraded substantially under corruptions, and generalization across datasets remains insufficient for real deployment.
Studying for a driving test using only smooth, empty roads. But the real test happens in heavy traffic and rain, so your skills might not hold up. In this paper, researchers test AI models on 'clean' LiDAR data versus data with artificial problems like fog. The smooth roads are the standard benchmarks, and the heavy traffic is the real-world corruptions and domain shifts they simulate.
Autonomous vehicles need to work in messy real-world conditions, not just in perfect labs. This research shows that current AI models might fail when it rains or in new cities, which could be dangerous. Better evaluation helps catch these issues before deployment.
The abstract doesn't say this work creates new segmentation models; it proposes an evaluation protocol. It tests only on specific hardware (Jetson AGX Orin) and eight corruption types. The protocol doesn't solve the robustness issues, just highlights them.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Toward Robust LiDAR Semantic Segmentation for Real-World Deployment: Evaluation under Coarse Labels, Adverse Conditions, and Domain Shifts — Samir Abou Haidar, Alexandre Chariot, Mehdi Darouich et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.