An End-to-End Collaborative Learning Approach for Connected Autonomous Vehicles in Occluded Scenarios

TL;DR

This work tackles safe navigation in occluded intersections where single-agent policies struggle. It introduces an end-to-end collaborative learning framework that uses V2V sharing of compressed LiDAR features and Multi-agent Proximal Policy Optimisation to train decentralized policies. The approach demonstrates superior safety and efficiency compared to independent RL and cooperative early fusion baselines while respecting practical bandwidth constraints, leveraging a CARLA-based occluded intersection environment. The findings highlight the benefits of collaborative perception for CAVs in occluded scenarios and underscore the importance of robust performance under realistic communication limits and sensor dropout. Overall, the paper advances practical, scalable MARL solutions for connected autonomous driving in challenging urban settings.

Abstract

Collaborative navigation becomes essential in situations of occluded scenarios in autonomous driving where independent driving policies are likely to lead to collisions. One promising approach to address this issue is through the use of Vehicle-to-Vehicle (V2V) networks that allow for the sharing of perception information with nearby agents, preventing catastrophic accidents. In this article, we propose a collaborative control method based on a V2V network for sharing compressed LiDAR features and employing Proximal Policy Optimisation to train safe and efficient navigation policies. Unlike previous approaches that rely on expert data (behaviour cloning), our proposed approach learns the multi-agent policies directly from experience in the occluded environment, while effectively meeting bandwidth limitations. The proposed method first prepossesses LiDAR point cloud data to obtain meaningful features through a convolutional neural network and then shares them with nearby CAVs to alert for potentially dangerous situations. To evaluate the proposed method, we developed an occluded intersection gym environment based on the CARLA autonomous driving simulator, allowing real-time data sharing among agents. Our experimental results demonstrate the consistent superiority of our collaborative control method over an independent reinforcement learning method and a cooperative early fusion method.

Figures (5)

• Figure 1: Collaborative CAV control. Some vehicles may be occluded by tall trees/buildings or other vehicles in the scene. With the use of LiDAR features from nearby agents, safe and efficient policies can be developed to avoid accidents.
• Figure 2: Collaborative MARL framework. LiDAR features are extracted from preprocessed point cloud data and shared within a V2V local communication network. The aggregated messages are then used to calculate the desired action through the actor network.
• Figure 3: Occluded scenarios. The first scenario corresponds to a blind intersection, where vehicles are occluded by tall trees. The second scenario corresponds to a blind summit.
• Figure 4: Learning curves for MAPPO with ground truth locations of occluded vehicles and MAPPO with collaborative perception.
• Figure 5: Collision and Success rate as a function of LiDAR point dropout.