3DGS Simulation + Human Experience + End-to-End RL

ParkingWorld

End-to-End Autonomous Parking Reinforcement Learning from Human Experience in 3DGS Simulation

Zhengcheng Yu, Changze Li, Haoran Liu, Tong Qin

Paper PDF Code

Simulation PSR

Real-Vehicle PSR

Standard Slots

What It Solves

Learning to park from corrected mistakes

ParkingWorld trains an end-to-end autonomous parking policy inside photorealistic 3D Gaussian Splatting parking scenes. Instead of relying only on ordinary RL rollouts or large expert datasets, the system couples what the autonomous policy did wrong with how humans corrected the failure. This correction-in-the-loop replay design makes training more sample-efficient, safer, and more transferable to real vehicles.

Photorealistic 3DGS scenes BEV perception-policy network Rollback correction replay Real Changan CS55 deployment

Framework

Closed-loop training inside reconstructed parking worlds

Interactive 3DGS autonomous parking simulator — Interactive ROS-based 3DGS simulator with rendering, dynamics, collision checking, rollback recording, and human-in-the-loop training.

CIL-SERL

Correction-in-the-loop sample-efficient RL

Human interventions are not simply mixed into a single demonstration buffer. ParkingWorld stores normal RL rollouts, human takeover segments, failed autonomous trajectories, and corrected rollback segments in coupled replay buffers. The policy therefore learns both the failure context and the corresponding recovery behavior.

Detect failure-prone autonomous states

Correct with human intervention and rollback

Replay paired failure and correction samples

Why 3DGS

Less domain gap, richer interaction

Comparison of CARLA LGSVL, real world, and 3DGS simulator training — ParkingWorld uses reconstructed 3DGS scenes to bridge simulator training and real-world deployment.

Simulation Experiment

Strong closed-loop performance on standard slots

Training and testing are conducted on five reconstructed 3DGS parking scenes containing 239 standard parking slots. Each method is evaluated over 200 trials.

Type	Method	PSR (%) ↑	PCR (%) ↓	PTR (%) ↓	PBR (%) ↓	NGS ↓
Rule-based	RS Curve	31.0	0.0	69.0	0.0	1.6
Rule-based	Hybrid A*	55.5	0.0	45.5	0.0	14.9
End-to-end	ParkingE2E	34.0	46.5	0.0	19.5	2.6
End-to-end	REAP-PPO	46.0	39.0	15.0	3.0	26.7
End-to-end	REAP-SAC	68.5	22.0	9.5	0.0	17.5
End-to-end	ParkingWorld	88.0	4.0	8.0	0.0	13.2

ParkingWorld trajectory visualization across different scenes

Real-World Experiment

Deployed on a Changan CS55

Four surround-view fisheye cameras stream images at 5 Hz to a dedicated computer, where ParkingWorld predicts target speed and steering commands. The commands are rolled out into a short-horizon kinematic trajectory and tracked by an onboard NUC using sampled NMPC at 50 Hz.

80.0%real PSR

5.0%near-collision rate

0.0%boundary crossing

Real vehicle deployment framework for ParkingWorld

Real-Vehicle Comparison

Robust transfer beyond simulation

Type	Method	PSR (%) ↑	PCR (%) ↓	PTR (%) ↓	PBR (%) ↓	NGS ↓
Rule-based	RS Curve	15.0	0.0	85.0	0.0	6.5
Rule-based	Hybrid A*	35.0	0.0	65.0	0.0	29.7
End-to-end	ParkingE2E	30.0	45.0	0.0	25.0	8.5
End-to-end	REAP-PPO	40.0	2.0	35.0	5.0	33.8
End-to-end	REAP-SAC	60.0	15.0	25.0	0.0	26.6
End-to-end	ParkingWorld	80.0	5.0	15.0	0.0	22.1

Citation

BibTeX

@article{parkingworld2026,
  title   = {ParkingWorld: End-to-End Autonomous Parking Reinforcement Learning from Human Experience in 3DGS Simulation},
  author  = {Yu, Zhengcheng and Li, Changze and Liu, Haoran and Qin, Tong},
  journal = {IEEE Robotics and Automation Letters},
  year    = {2026}
}