NeuralParker: A Reinforcement Learning Planner for Irregular Parking Environments
Authors: Zihan Wang, Bai Huang, Yang Guan, Xiao Li, Haoyu Xu, Naizheng Wang, Shengbo Eben Li
Organizations: School of Life Sciences, Tsinghua University, Beijing 100084, China · School of Statistics and Mathematics, Central University of Finance and Economics, Beijing 100081, China · School of Vehicle and Mobility, Tsinghua University, Beijing 100084, China · Meituan Autonomous Delivery, Beijing 100102, China
Automated parking commonly assumes marked slots and short approach maneuvers. Delivery and service vehicles, however, may need to reach an operator-specified pose in an irregular bounded environment from a distant start. Existing learning-based parking planners often rely on local observations, which can restrict long-range route reasoning. To address this problem, we present NeuralParker, a reinforcement learning-based hybrid planner for arbitrary-pose parking. NeuralParker encodes full-environment obstacle and boundary geometry in a target-relative vertex representation, allowing the policy to retain route-defining context throughout the approach. It further couples a learned curvature--length arc policy with an in-loop terminal ensemble that selects from diverse cubic Hermite connections using a curvature-regularized cost. We also establish factorial and long-range route-choice benchmarks to evaluate planning success and trajectory quality. Experiments on these benchmarks show that NeuralParker achieves higher planning success and better overall trajectory quality than the evaluated baselines, while ablation studies support the benefits of the target-relative global representation and terminal ensemble. Finally, a real-vehicle evaluation confirms that the planner transfers effectively to real delivery-vehicle perception at a working parking site, planning successfully at low computational cost.
Figures & tables
Fig. 1: Overview of the NeuralParker architecture. The observation is decomposed into the ego pose, obstacle vertices, boundary segments, and local LiDAR signals. The scene encoder embeds the four streams, uses self-attention to aggregate the obstacle and boundary tokens, and fuses the resulting global geometry with the ego and LiDAR features. The shared representation is passed to the actor and critic networks, which output the curvature–length arc (κ,l) and the state-value estimate, respectively.
Fig. 2: Representative held-out scenes from the Factorial Parking Benchmark (left) and the Topology-Stress Benchmark (right). The Factorial block arranges start-separation regimes by row and target-pose types by column; the Topology-Stress block arranges route-choice families by row and difficulty by column. All panels preserve equal x – y metric scaling. Red T markers denote target poses and headings, and axes are in meters.
Fig. 3: NeuralParker improves success rate and trajectory quality over the adapted HOPE planner on the Factorial Parking Benchmark. Success uses all 1080 episodes per seed; trajectory metrics use the pairwise common-success cohort ( n=773/748/790 ). Bars and error bars show the three-seed mean ± sample SD. NP denotes NeuralParker.
Method
Succ. ↑
L (m) ↓
Rev. ↓
Δκ↓
Unified NP
0.974 ± 0.007
19.77 ± 0.41
1.02 ± 0.10
1.77 ± 0.10
Hybrid A* → NP
0.969 ± 0.002
22.22 ± 0.50
1.42 ± 0.05
3.05 ± 0.09
Hybrid A* → HOPE
0.686 ± 0.014
36.36 ± 0.38
6.97 ± 0.34
5.16 ± 0.10
TABLE I: Unified Versus Staged Planning. Succ. uses all 1080 episodes per seed; path metrics use the all-method common-success cohort ( n=700/719/726 ). For staged rows, Δκ sums the within-prefix and downstream variations and excludes the single handoff curvature jump.
Fig. 4: Unified planning keeps one approach decision where a staged handoff splits it. Each planner is drawn as one line, so the staged arm’s arcs and terminal connector share the dashed green line. One checkpoint drives both learned arms on this pinned pocket case; the dashed circle is the 10 m handoff boundary.
Added representation
Succ. ↑
L (m) ↓
Rev. ↓
Δκ↓
NeuralParker backbone
none
0.897 ± 0.010
37.84 ± 4.01
2.96 ± 0.58
5.74 ± 0.69
BEV
0.848 ± 0.030
38.72 ± 1.07
4.41 ± 0.27
6.25 ± 0.35
vertices
0.916 ± 0.023
35.40 ± 0.93
2.77 ± 0.10
5.06 ± 0.55
HOPE-style backbone
none
0.690 ± 0.018
58.53 ± 1.54
21.70 ± 1.03
13.34 ± 0.63
TABLE II: Scene-Representation Ablation. Succ. uses all 720 episodes per seed; path metrics use the six-method common-success cohort ( n=328/317/281 ).
Fig. 5: Absolute family-level success rates show that representation effects depend on route-choice family and policy backbone on the Topology-Stress Benchmark. Each family contains 240 episodes per seed. Every group compares local-only, BEV, and vertices; bars and error bars show the three-seed mean ± sample SD, and open circles show individual seeds.
Fig. 6: Both terminal rules are feasible from the marked handoff pose, but the ensemble selects a different terminal shape. Replaying only the connector from a shared handoff isolates candidate selection from the policy prefix, which is shown for context.
Terminal rule
Succ. ↑
L (m) ↓
Rev. ↓
Δκ↓
Fixed Hermite
0.965 ± 0.007
20.13 ± 0.69
1.37 ± 0.09
2.36 ± 0.15
Ensemble, eval. only
0.972 ± 0.006
18.76 ± 0.23
1.11 ± 0.06
1.91 ± 0.07
Ensemble, train + eval.
0.974 ± 0.007
18.69 ± 0.11
1.15 ± 0.06
1.86 ± 0.08
TABLE III: Terminal-Connector Ablation on the Factorial Parking Benchmark. Succ. uses all 1080 episodes per seed; path metrics use the three-configuration common-success cohort ( n=1003/1000/1015 ).
Start curriculum
Succ. ↑
L (m) ↓
Rev. ↓
Δκ↓
No curriculum
0.969 ± 0.015
19.90 ± 0.75
1.43 ± 0.10
2.23 ± 0.12
Geodesic curriculum
0.974 ± 0.007
18.76 ± 0.16
1.17 ± 0.05
1.90 ± 0.06
TABLE IV: Start-Curriculum Ablation. Succ. uses all 1080 episodes per seed; path metrics use the pairwise common-success cohort ( n=1043/1005/1028 ). Values are three-seed mean ± sample SD.
Fig. 7: Real-vehicle experiments with a simpler NeuralParker configuration. (a)–(d) operating site, perception visualization, camera view, and site-specific simulation reconstruction; (e)–(h) two simulation-validated cases comparing Hybrid A* with NeuralParker-Guided A*.
Method
Succ. ↑
L (m) ↓
Rev. ↓
Plan time (ms) ↓
NeuralParker
0.60 (0.93)
7.0
1.20
0.37
NP-Guided A*
0.72
7.7
1.17
4.24
Hybrid A*
0.67
7.2
1.13
5.26
TABLE V: Real-Vehicle Experiment on 47 Difficult Clips. Succ. uses all clips; L and Rev. are computed on each method’s own successful clips.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Setting
Coordinate frame
Target at (0,0) with heading +y
Obstacle capacity
20 slots × 16 values
Boundary capacity
10 slots × 4 values
Local LiDAR
120 rectified footprint-clearance beams, 360∘ , 10 m ray cap
Geometry tokens
16-D, one attention head
Global–local fusion
128-D, depth 1, four 32-D heads
Appendix
TABLE VI: NeuralParker Configuration
Parameter
Setting
Learning rate / schedule
3×10−4 / linear decay
Discount γ / GAE λ
0.99 / 0.95
PPO clip / value coefficient
0.15 / 0.25
Entropy coefficient / gradient norm
0.005 / 0.5
Advantage / reward normalization
yes / yes
Epochs / steps per epoch
240 / 60,000
Appendix
TABLE VII: NeuralParker Training Configuration
Fig. 8: Learned guidance reduces Hybrid A* search time mostly on the harder real-vehicle clips. The 47 held-out clips are ordered by Hybrid A* search time, which serves as a difficulty proxy, and both planners’ times are shown in that order.
Robotics and Autonomous Systems Thrust, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou 511453, China · Cheng Kar-Shun Robotics Institute, The Hong Kong University of Science and Technology, Hong Kong SAR, China