无人驾驶CoRL 20252025

DrivePolicy: Learning Driving Policies from Human Demonstrations

Soheil HumanWayve, University of Cambridge

摘要

We present DrivePolicy, a learning framework that acquires driving policies from human demonstrations. DrivePolicy uses behavior cloning combined with reinforcement learning to learn robust driving policies that handle complex urban scenarios including intersections, roundabouts, and highway merging.

driving policyhuman demonstrationbehavior cloningreinforcement learningurban scenario

技术细节

数据集

nuScenes Waymo Open DatasetHuman Driving Data

仿真平台

CARLA nuScenes Devkit

模型骨架

Imitation Learning + BEV Transformer

编码器

BEV Encoder + Ego State Encoder

解码器

Planning Decoder (6 layers)