REAP: Reinforcement-Learning End-to-End Autonomous Parking with Gaussian Splatting Simulator for Real2Sim2Real Transfer

Researchers introduce REAP, a reinforcement learning-based autonomous parking system that uses Gaussian Splatting to simulate real-world environments for training, then transfers the model to physical vehicles. The method addresses limitations of traditional multi-stage parking approaches by jointly optimizing perception and planning, achieving successful parking in extreme scenarios like mechanical slots.

REAP represents a meaningful advancement in autonomous vehicle technology by tackling a persistent challenge in self-driving systems: reliable parking in constrained spaces. Traditional parking approaches fragment the problem into separate perception and planning stages, causing errors to compound through the pipeline. This research consolidates those stages into an end-to-end system, enabling holistic optimization that better handles edge cases where conventional methods fail. The integration of Soft Actor-Critic reinforcement learning with behavior cloning from rule-based planners accelerates convergence while maintaining practical performance, addressing a core inefficiency in RL-based approaches. The 3D Gaussian Splatting simulator bridges a critical gap in autonomous vehicle development: the sim-to-real transfer problem. By reconstructing real parking environments as photorealistic digital scenes, the system trains models in simulation while maintaining sufficient fidelity for direct deployment on physical vehicles. This reduces expensive real-world testing iterations and enables safer algorithm development. For the autonomous vehicle industry, successful parking solutions directly impact user experience and market adoption, as parking remains one of the most failure-prone autonomous driving tasks. The demonstrated capability in mechanical slots—notoriously difficult scenarios—suggests meaningful progress toward production-ready systems. The soft collision penalty mechanism's approach to safety during learning could inform how other safety-critical autonomous tasks are trained. However, the research remains academic; practical deployment at scale requires validation across diverse real-world conditions, weather variations, and edge cases beyond the tested parking scenarios.

• →REAP combines end-to-end reinforcement learning with behavior cloning to improve parking success in extreme scenarios like mechanical slots
• →3D Gaussian Splatting enables photorealistic sim-to-real transfer, reducing gap between simulation training and physical vehicle deployment
• →Soft Actor-Critic framework with asymmetric learning improves training efficiency over traditional RL approaches for autonomous parking
• →Joint optimization of perception and planning eliminates error accumulation from multi-stage approaches
• →Soft predictive collision penalties reduce unsafe obstacle-approach actions during reinforcement learning training