Kobbi, Islem (INRIA), ATOUI, Hussam (Valeo), Nashashibi, Fawzi (INRIA)

Reinforcement Learning for Mid-To-Mid Motion Planning in Autonomous Driving

Scheduled for presentation during the Regular Session "S26c-Motion Planning, Trajectory Optimization, and Control for Autonomous Vehicles" (TH-LA-T26), Thursday, November 20, 2025, 16:40−17:00, Broadbeach 1&2

2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC), November 18-21, 2025, Gold Coast, Australia

This information is tentative and subject to change. Compiled on October 18, 2025

Abstract

This paper proposes a reinforcement learning (RL) based mid-to-mid motion planning framework for autonomous vehicles, where an agent receives structured mid-level observations and outputs continuous paths suitable for tracking. To accelerate training and avoid the inefficient early phase of RL, the agent is first initialized using imitation learning (IL) via behavior cloning (BC) from a rule-based motion planner (RBMP). RL training then takes over using Proximal Policy Optimization (PPO), where the agent learns directly from interaction to achieve high-performance driving behavior. This mid-to-mid architecture preserves the transparency and modularity of classical decision-making pipelines while leveraging the full capabilities of RL. Experimental results show that the RL-trained agent not only surpasses the initial IL-based policy by a wide margin but also clearly outperforms the RBMP baseline, demonstrating that RL is solely responsible for achieving superior, robust, and adaptable motion planning.