Learning to Talk: Emergent CAM/CPM Scheduling via MAPPO in Constrained V2X Networks
Résumé
Broadcasting every Cooperative Awareness Message (CAM) and Collective Perception Message (CPM) without prioritisation can saturate the NR-V2X sidelink that connected and automated vehicles rely on for safety-critical exchanges. We model joint CAM/CPM scheduling as a cooperative partially observable Markov decision process (C-POMDP) and learn a unified policy using Multi-Agent Proximal Policy Optimisation (MAPPO). A tunable reward-controlled by two weights (α,β)-lets the policy traverse the full information-versus-bandwidth Pareto frontier without redesign. In this formulation, "knowledge" is defined as the cumulative number of dynamic objects whose state is simultaneously known to all road users. The baseline is a naïve scheme that broadcasts every CAM at 10 Hz and every CPM as soon as it is generated. In a SUMO-based simulation of urban Toulouse, enriched with line-of-sight propagation and probabilistic vision detection, the learned scheduler reduces channel load by 43.2% while preserving 82.1% of baseline scene knowledge at α = 3.0, β = 2 × 10 -4 . This results in a 42.9% improvement in knowledge per byte transmitted. Qualitative analysis reveals emergent role specialisation and bursty, context-aware message scheduling that static heuristics fail to capture. These findings highlight the potential of collaborative policy learning to meet strict bandwidth constraints and adaptively balance CAM and CPM traffic-without relying on hand-tuned rules. Code and datasets are publicly released for full reproducibility.
Mots clés
- Intelligent Transportation Systems
- Proximal Policy Optimisation
- Multi-Agent Reinforcement Learning
- V2X communication
- Collective Perception Messages (CPM)
- Cooperative Awareness Messages (CAM)
- Cooperative Awareness Messages (CAM) Collective Perception Messages (CPM) V2X communication Multi-Agent Reinforcement Learning Proximal Policy Optimisation Intelligent Transportation Systems
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |