REAP: End-to-End Autonomous Parking via Reinforcement Learning and Gaussian Splatting
Researchers have introduced REAP, a novel reinforcement learning-based method for end-to-end autonomous parking, designed to overcome challenges in extreme scenarios like narrow mechanical and dead-end slots. Traditional multi-stage parking systems often fail in these conditions due to error accumulation, while existing end-to-end approaches struggle with data costs or inefficient exploration. REAP addresses these issues by employing Soft Actor-Critic (SAC) within an asymmetric framework to enhance training efficiency and inference performance. The model accelerates convergence by distilling rule-based planner capabilities through behavior cloning and incorporates a soft predictive collision penalty to minimize accidents. A key innovation is the Real2Sim2Real simulator, which uses 3D Gaussian Splatting (3DGS) to create high-fidelity digital twins of real-world scenes for training. This approach bridges the simulation-to-reality gap, allowing the trained network to deploy directly onto physical vehicles. Experimental results demonstrate that REAP successfully executes parking maneuvers in various complex environments, proving the feasibility of end-to-end reinforcement learning for highly constrained parking tasks.
Wire timeline
REAP: End-to-End Autonomous Parking via Reinforcement Learning and Gaussian Splatting
Researchers have introduced REAP, a novel reinforcement learning-based method for end-to-end autonomous parking, designed to overcome challenges in extreme scenarios like narrow mechanical and dead-end slots. Traditional multi-stage parking systems often fail in these conditions due to error accumulation, while existing end-to-end approaches struggle with data costs or inefficient exploration. REAP addresses these issues by employing Soft Actor-Critic (SAC) within an asymmetric framework to enhance training efficiency and inference performance. The model accelerates convergence by distilling rule-based planner capabilities through behavior cloning and incorporates a soft predictive collision penalty to minimize accidents. A key innovation is the Real2Sim2Real simulator, which uses 3D Gaussian Splatting (3DGS) to create high-fidelity digital twins of real-world scenes for training. This approach bridges the simulation-to-reality gap, allowing the trained network to deploy directly onto physical vehicles. Experimental results demonstrate that REAP successfully executes parking maneuvers in various complex environments, proving the feasibility of end-to-end reinforcement learning for highly constrained parking tasks.
cs.AI updates on arXiv.org