MILER: Semantic Mid-Level Representation for Sim-to-Real Reinforcement Learning in Unstructured Autonomous Driving

Summary

The paper presents MILER, a reinforcement-learning framework for autonomous driving in unstructured environments built around zero-shot sim-to-real transfer. Training takes place entirely in a custom GPU-accelerated simulator based on a semantic mid-level representation (MLR): maps are synthesized from Perlin noise into four semantic classes, vehicle dynamics use a bicycle model, and the policy is trained with PPO using a discrete action head (jerk and steering-rate change), an undiscounted objective (gamma=1), curriculum learning, and state normalization. At deployment, an adapted BEVFusion converts camera and LiDAR data into semantic BEVs matching the simulator's representation, and a trajectory-alignment strategy lets a virtual vehicle inside the simulator execute the policy's actions while the real vehicle tracks it, decoupling simulated from real dynamics. The authors evaluate on a 3.0 km test track with two vehicles, driving 17.3 km over three laps each without human intervention, reaching up to 33.6 km/h, with the full stack on a Jetson AGX Orin (1.9 ms inference). Ablations show that identical HD-map BEVs narrow the inter-vehicle behavioral gap (mean trajectory distance 0.078 m vs 0.124 m) and that removing trajectory alignment causes oscillations and failure. Limitations include vehicle B stopping under rain and darkness, absent human-like swing-out maneuvers, a speed ceiling imposed by BEV range, static obstacles only, and evaluation on a single private track without repeated trials, quantitative baselines, or code release.