无人驾驶CVPR 20252025

LiDAR-LLM: Language Model for LiDAR Point Cloud Understanding

The Cloud Collaboration et al.Zhejiang University, Horizon Robotics

摘要

We present LiDAR-LLM, a language model that understands LiDAR point clouds for autonomous driving. LiDAR-LLM processes raw point cloud data and generates natural language descriptions of the 3D environment, enabling interpretable perception and reasoning about driving scenes.

LiDAR understandinglanguage modelpoint cloud3D perceptioninterpretable perception

技术细节

数据集

SemanticKITTI nuScenes Waymo Open Dataset

测试机器人

Velodyne VLP-16Hesai Pandar64

模型骨架

LLaMA-2 7B + LiDAR Projector

编码器

PointNet++ + LiDAR Feature Projector

解码器

LLaMA-2 7B Language Decoder (32 layers)