Intensity for Any Scenes/Simulators
Add synthesized intensity to calibrated point clouds, without per-scene optimization in the evaluated settings.
NIDAR is a plug-and-play LiDAR intensity synthesis module that augments arbitrary LiDAR simulators and re-simulation pipelines. It infers physically consistent intensity from low-cost RGB inputs and geometry, requiring no per-scene training/optimization and no intensity ground truth during deployment.
As a crucial signal for perception and state estimation, LiDAR intensity provides complementary reflectance cues beyond geometry. Despite its importance, most existing simulation pipelines either omit intensity altogether or rely on reconstruction-based approaches that require scene-specific real intensity supervision. Such dependence on real groundtruth data not only incurs substantial data acquisition costs but also hinders scalability and limits generalization across scenes, sensors, and simulation platforms.
To address this gap, we propose NIDAR, a general and scene-agnostic intensity synthesis module that augments arbitrary LiDAR simulators and re-simulation pipelines. NIDAR infers physically consistent intensity from low-cost RGB inputs and geometry, requiring no per-scene training/optimization and no intensity ground truth during deployment. We implement NIDAR as a hierarchical UNet+DeepLabV3 intrinsic decomposition network with lightweight distribution calibration, achieving real-time feed-forward inference while avoiding the hours-level scene fitting cost of reconstruction-based baselines. Comprehensive experiments on nuScenes and Waymo verify the effectiveness and cross-scene generalization of NIDAR. We further validate its scalability by integrating it into Isaac Sim, UE5, and generative simulation pipelines. In addition, downstream SLAM evaluations in Isaac Sim demonstrate that NIDAR-generated intensity improves localization accuracy and significantly reduces drift, especially in challenging scenarios. Our approach provides a unified solution that equips existing LiDAR simulation systems with realistic intensity generation capability. Source code is available on GitHub.
NIDAR pipeline. Cross-spectral pseudo-NIR prior, intrinsic decomposition (reflectance/shading) with mask-aware intensity supervision, and distribution calibration for plug-and-play LiDAR intensity synthesis.
We compare to reconstruction‑based methods; point clouds are colorized by synthesized intensity to highlight visual differences. Importantly, NIDAR was not trained on the nuScenes dataset — these results demonstrate strong zero‑shot generalization. In these scenes, the baseline methods' point-cloud intensity differ notably from the GT, while NIDAR's intensity match the GT most closely.
To validate the practical utility of NIDAR-generated intensity, we evaluate its impact on downstream SLAM tasks in Isaac Sim. The results demonstrate that incorporating NIDAR intensity significantly improves localization accuracy and reduces trajectory drift compared to geometry-only approaches.
We integrate NIDAR with UE5 to add intensity information to simulated point clouds.
Simulator: MuJoCo · Policy: BC-RNN (recurrent behavior cloning).
We explore NIDAR’s potential in imitation learning: a robot uses RGB, depth and NIDAR-derived visual cues to grasp and lift a target cube among distractors. The NIDAR output is derived from RGB, without an additional NIR sensor. Each clip records a trained BC-RNN policy predicting actions from successive observations in a closed loop.
Selected qualitative rollout excerpts. Outcomes refer to the shown interval, not aggregate success rates.
Three red cubes, one blue cube and one green cube share the tabletop. The task is to grasp and lift the central red cube with a wood-grain texture.
RGB + depth
RGBD + grayscale cues
RGBD + NIDAR output
@misc{zhang2026nidar,
author = {Junjie Zhang and Jie Yin and Kefei Qian and Jie Li and
Mengpei Jia and Yajuan Dun and Wenbo Chu and Guofa Li},
title = {{NIDAR}: {NIR}-Guided Intrinsic Decomposition for Scalable Scene-Agnostic {LiDAR} Intensity Reconstruction},
year = {2026},
eprint = {2609.36878},
archivePrefix = {arXiv},
primaryClass = {cs.RO},
url = {https://arxiv.org/abs/2609.36878},
}