Reproduce “ParkingWorld: End-to-End Autonomous Parking Reinforcement Learning from Corrective Experience in 3DGS Simulation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhengcheng Yu, Changze Li, Haoran Liu, Tong Qin
The ParkingWorld framework introduces an end-to-end autonomous parking system, leveraging a photorealistic 3D Gaussian Splatting (3DGS) simulator and a novel Correction-in-the-Loop Sample-Efficient Reinforcement Learning (CIL-SERL) method. This approach attained an 88.0% parking success rate in simulation and an 80.0% success rate on a real vehicle, demonstrating robust sim-to-real transfer and a 5.0% near-collision rate in real-world tests.
Key insights
- Photorealistic 3DGS simulation can effectively bridge the sim-to-real gap for autonomous parking, enabling policies trained purely in simulation to generalize to real vehicles without fine-tuning.
- Learning from human-corrected mistakes, particularly through explicitly paired failure-correction trajectories, dramatically improves reinforcement learning's sample efficiency, safety, and robustness.
- Structuring the replay memory with a multi-level buffer, which prioritizes sampling failed autonomous segments followed by successful human corrections, optimizes the learning signal for complex, constrained tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.