PaperlayerView exact revision on arXiv

Source-linked AI summary

NeuralParker: A Reinforcement Learning Planner for Irregular Parking Environments

Zihan Wang, Bai Huang, Yang Guan, Xiao Li, Haoyu Xu, Naizheng Wang, Shengbo Eben Li

arXiv:2608.24485v1cs.ROcs.LG

TL;DR

NeuralParker addresses arbitrary-pose parking in irregular bounded environments, where local observations can limit long-range route reasoning. It uses target-relative full-environment geometry with a learned arc policy and curvature-regularized Hermite terminal ensemble, and outperforms evaluated baselines on success and trajectory quality.

  • Problem

    Delivery and service vehicles may need to reach distant operator-specified poses in irregular environments, while existing learned planners often rely on local observations that omit target-side geometry.

    Source: §I · p. 1

  • Method

    NeuralParker combines target-relative obstacle and boundary vertices, local clearance rays, a learned curvature–length arc policy, and an in-loop Hermite terminal ensemble.

    Source: §I · p. 2, §VI · pp. 8–9

  • Results

    NeuralParker achieves the highest success rate on factorial and long-range route-choice benchmarks, with shorter and smoother trajectories than compared planners.

    Source: §I · p. 2

  • Takeaways & Limitations

    Target-relative vertices provide the strongest aggregate success and trajectory quality under both evaluated policy backbones, while the terminal ensemble improves these outcomes with or without policy retraining.

    Source: §I · p. 2, §VI · p. 9

Abstract

from arXiv · show

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.

I. Introduction

NeuralParker targets arbitrary-pose parking, where vehicles must approach operator-specified poses through distant, irregular, bounded environments. It combines target-relative global geometry, learned arcs, and an in-loop terminal ensemble, evaluated with route-choice benchmarks.

  • Arbitrary-pose parking requires reaching a task-defined pose beside operational locations from distant starts through irregular bounded environments.Source: §I · p. 1
  • Existing learned planners may omit target-side geometry before selecting a long-range approach because they use compact, ego-centered, or finite-range observations.Source: §I · p. 1
  • An attention-based policy outputs curvature–length arcs, while an in-loop cubic-Hermite ensemble tests diverse terminal hypotheses under feasibility and curvature constraints.Source: §I · p. 2
  • NeuralParker represents full-environment obstacle and boundary geometry in a target-relative vertex representation and augments it with local clearance rays.Source: §I · pp. 1–2
  • Factorial and long-range route-choice benchmarks separate full-set success from common-success trajectory quality, with NeuralParker reporting the highest success and shortest, smoothest trajectories among compared planners.Source: §I · p. 2

A. Traditional Parking Planners

Traditional and learning-based parking planners differ in how explicitly they retain vehicle geometry, environment geometry, and analytic planning stages. NeuralParker is positioned as a learned policy that preserves map-wide context and checked feasibility.

Traditional planners

NeuralParker

Learning-based planners

Hybrid planners

B. Target-Relative Full-Environment Observation

NeuralParker expresses scene geometry in a target-relative metric frame while retaining near-field clearance information. This preserves target-side context across ego–target distances and applies explicit capacity and collision conventions.

C. Arc Policy and Scene Encoder

NeuralParker uses a scene encoder and actor–critic policy to generate curvature–length arcs, then applies a gated Hermite terminal ensemble. Candidate connections are filtered for geometry and curvature feasibility before selection.

Arc policy

Scene encoder

Terminal ensemble

E. Reward and Training

NeuralParker trains with PPO using an obstacle-aware reverse start curriculum and an in-loop Hermite ensemble whose feasibility decisions are included in rollout transitions.

Reward

  • The reported path-quality reward penalizes near-zero-length actions and associated sign jitter, while tracking curvature-sign changes across policy arcs.For the first arc, history-dependent terms are zero.Source: E. Reward and Training · p. 4
  • On successful episodes, terminal rewards incorporate the selected Hermite segment’s length and discrete curvature variation; collision or boundary failure receives −2501.Direct terminal states omit Hermite terms, and timeouts receive no success bonus.Source: E. Reward and Training · p. 4

Training

  • An obstacle-aware reverse curriculum orders initial states by 1.5 m-grid geodesic distance, beginning at 5% of the available range and expanding during training.The matched adapted-HOPE control is trained without a start-state curriculum.Source: E. Reward and Training · pp. 4–5
  • Training samples scenes and curriculum starts, rolls out policy curvature–length actions, evaluates terminal candidates, stores transitions, and updates actor and value parameters with clipped PPO.The same Hermite feasibility test determines successful transitions in rollout data, rather than serving only as post-processing.Source: E. Reward and Training · p. 5

A. Scenario Design

The evaluation combines factorial coverage with a topology-stress benchmark for long-range route choices, using frozen full-range starts and common-success trajectory metrics.

Scenario Design

  • The Factorial Parking Benchmark has 81 training scenes and 27 structurally disjoint test scenes across three start-distance regimes, three parking types, and three difficulty levels.Each evaluation seed runs 40 starts per test scene, totaling 1080 episodes.Source: A. Scenario Design · p. 5
  • The Topology-Stress Benchmark contains 48 training scenes and 18 structurally disjoint test scenes crossing three route-choice families with medium and hard difficulty.Its target-side structure initially lies beyond the 10 m local LiDAR view, with 720 evaluation episodes per seed.Source: A. Scenario Design · p. 5

Baselines

  • HOPE is the principal learned baseline, adapted to NeuralParker’s collision, boundary, candidate-checking, and fixed-Hermite conventions for fair planner-level comparison.The adapted baseline retains HOPE’s action mask and Reeds–Shepp controller.Source: B. Baselines · p. 5
  • The real-vehicle study compares Hybrid A* with NeuralParker-Guided A*, which preserves complete search while biasing expansion order toward a learned reference path.The underlying search and feasibility tests remain unchanged.Source: B. Baselines · p. 5

Metrics and Training

  • Evaluation reports full-set success rate, path length, full-trajectory reversals, and cumulative curvature change, with trajectory metrics computed from common successful episodes.The protocol uses a fixed full-range start set replayed across methods and training seeds.Source: C. Metrics and Training · pp. 5–6

V. Results

Results compare complete planners, unified and staged decompositions, and representative benchmark scenes using full-set success alongside common-success trajectory metrics.

Planner Comparison

  • NeuralParker improves success rate and trajectory quality over adapted HOPE on the Factorial Parking Benchmark.Success uses all 1080 episodes per seed, while trajectory metrics use pairwise common-success cohorts of 773/748/790.Source: Figure 3 · p. 6, A. Comparison With Baseline · p. 6

Unified Versus Staged Planning

  • Unified planning achieves the lowest path costs in comparison with two staged decompositions using a Hybrid A* prefix and a fixed 10 m handoff.The shared staged prefix uses continuous-curvature smoothing when feasible and a frozen raw-path fallback.Source: A. Comparison With Baseline · pp. 6–7
  • Handoff to the same NeuralParker policy retains comparable success but degrades all three trajectory metrics, whereas handoff to adapted HOPE also reduces success.Neither downstream policy is retrained for the handoff.Source: A. Comparison With Baseline · p. 7

B. Global Geometry Ablation

The global-geometry ablation tests whether target-relative scene representations improve route-choice planning across route families and policy backbones. The terminal-connector ablation separately tests whether a curvature-aware candidate ensemble improves terminal selection and trajectory quality.

Global Geometry

  • The Topology-Stress Benchmark places the target-side opening outside the local view, requiring earlier selection of the approach side.This setting is designed to expose the value of explicit global geometry for route choice.Source: B. Global Geometry Ablation · p. 7
  • Vertices produce small gains on Types 1 and 3 and little change on Type 2 with the NeuralParker backbone, while BEV remains below local-only in all families.The family-level comparison uses three-seed means with sample standard deviations and 240 episodes per family per seed.Source: B. Global Geometry Ablation · p. 7, Figure 5 · p. 8
  • With the HOPE-style backbone, both added representations improve Types 2 and 3 but not Type 1, with appreciable between-seed variation in several variants.These distinct family responses motivate retaining Types 1 and 3 as separate route-choice families.Source: B. Global Geometry Ablation · p. 7

Terminal Connector

  • Replacing the fixed nine-candidate Hermite connector with the ensemble at evaluation improves success rate and all three path-quality metrics without changing the policy.The ablation isolates the immediate effect of enlarging the terminal candidate set.Source: C. Terminal Connector Ablation · p. 7, Table III · p. 8
  • Training with the same ensemble yields the highest success rate, shortest paths, and lowest curvature variation, while reversal count matches evaluation-only replacement within one seed standard deviation.The ensemble varies terminal hypotheses and endpoint tangent scales, rejects infeasible candidates, and selects a low-cost feasible connection.Source: C. Terminal Connector Ablation · p. 8
  • Both fixed and ensemble connectors are feasible from the same policy handoff, but the ensemble selects a different terminal shape.Replaying only the connector isolates candidate selection from the learned policy prefix.Source: C. Terminal Connector Ablation · p. 8, Figure 6 · p. 8

D. Start-Curriculum Ablation

The study compares start-ordering choices and evaluates NeuralParker in real-vehicle-derived scenarios. Across the reported evaluations, NeuralParker retains strong planning success and trajectory-quality results, with remaining scope boundaries in the benchmark and validation setup.

Start-Curriculum Ablation

  • Geodesic start ordering achieves similar fixed full-range success to no curriculum while improving all three common-success trajectory metrics.The comparison shares representation, reward, network, terminal connector, training budget, and evaluation starts, isolating start ordering.Source: D. Start-Curriculum Ablation · p. 8

Real-Vehicle Evaluation

  • The real-vehicle evaluation uses perception recorded by operating delivery vehicles, an HD map, and an assigned parking pose, with paths executed and validated in simulation.The evaluated configuration shares the target-relative geometry representation and arc policy, uses a fixed nine-candidate Hermite connector, and omits policy-level LiDAR.Source: E. Real-Vehicle Experiments · p. 8, §VI · p. 9
  • NeuralParker-Guided A* finds indirect maneuvers and reduces expanded nodes when both planners agree on the maneuver.The representative cases are shown in Fig. 7(e)–(h).Source: E. Real-Vehicle Experiments · p. 8, Figure 7 · p. 9

Benchmark Results

  • NeuralParker attains the highest success rate on both factorial and long-range route-choice benchmarks, with the shortest and smoothest trajectories among compared planners.These benchmarks separate full-set success from common-success trajectory quality.Source: §VI · p. 9

Scope Boundary

  • The reported results use reference-point collision checks, static procedurally constructed 2-D benchmarks, fixed object budgets, and three policy-training seeds.Dynamic actors, perception uncertainty, broader object counts and topologies, and closed-loop validation of the vertices-plus-LiDAR planner remain future work.Source: §VI · p. 9

Appendix A NeuralParker Implementation Details

The implementation specifies NeuralParker’s observation, network, action, terminal-connector, and training configurations for reproducible simulation experiments. The design combines target-relative geometry and rectified local LiDAR with attention-based policy processing and Hermite terminal handling.

Observation and Network

  • Unused geometry slots use the −104 sentinel, and rectified clearance is s_l_i = max(d_i−b_i, 0) without observation normalization.Circular obstacles use 16-segment rings for ray casting, with the same benchmark caster used during training and evaluation.Source: A. Policy and Planner Configuration · p. 9

Learned Action and Terminal Connector

  • Learned arcs use 30 sampled reference points for hard collision and boundary checks, while Hermite candidates use 100 samples.The same feasibility and sampling conventions support the reported simulation pipeline.Source: A. Policy and Planner Configuration · p. 9

Training Configuration

  • Training checkpoints are selected by per-epoch mean reward over 32 sampled full-range episodes, without using frozen evaluation episodes.Snapshots are saved every 20 epochs, and selected epochs differ across curriculum-enabled and no-curriculum seeds.Source: B. Training Configuration · p. 9, Table VII · p. 10

Observation and Network

  • NeuralParker uses target-relative obstacle and boundary geometry with 120 rectified footprint-clearance LiDAR beams capped at 10 m.The configuration allocates 20 obstacle slots with 16 values each and 10 boundary slots with 4 values each.Source: Appendix A · p. 9, B. Training Configuration · p. 10
  • Geometry tokens are 16-D with one attention head, while global–local fusion is 128-D with depth 1 and four 32-D heads.Separate actor and critic MLPs use [64, 128, 256, 64] networks.Source: B. Training Configuration · p. 10

Baseline Configuration

  • The adapted HOPE baseline preserves its actor–critic architecture, PPO implementation, reward, and action parameterization while using full-range starts without a curriculum.It retains a native 42-anchor LiDAR action mask, a Reeds–Shepp controller, and a fixed nine-candidate Hermite terminator.Source: A. Policy and Training Contract · p. 10, Appendix B · p. 10

B. Benchmark Interface and Analytic Control

The benchmark interface standardizes collision and boundary handling while comparing controlled scene representations and analytic connectors. Evaluation also measures how learned guidance changes Hybrid A* search effort on difficult real-world-derived clips.

Analytic Control

  • Adapted HOPE uses hard reference-point collision and boundary termination, replacing its collision-and-retreat behavior.It retains signed footprint clearance d_i−b_i with running mean-and-standard-deviation normalization, unlike NeuralParker’s rectification.Source: B. Benchmark Interface and Analytic… · p. 10
  • HOPE’s learned-action curvature bound is tan(0.75)/2.8 ≈ 0.333 m−1, while the attached fixed-Hermite terminator uses a common 0.68 m−1 feasibility cap.RS candidates are generated and checked at 0.1 m spacing; learned arcs and accepted RS chunks use 30 reference points, and Hermite candidates use 100.Source: B. Benchmark Interface and Analytic… · p. 10

Controlled Representation Inputs

  • The controlled representation study compares none, a BEV raster, and target-relative vertices over a shared target-and-signed-LiDAR base.Each configuration retains HOPE’s PPO, reward, and action parameterization, uses the geodesic start curriculum, and disables the action mask and RS controller.Source: C. Controlled Representation Inputs · p. 10
  • BEV controls use an ego-aligned 64 × 64 RGB raster and a frozen convolutional autoencoder pretrained on 50,000 tail-enabled images.The NeuralParker-backbone control disables the trajectory tail, while the HOPE-style control retains up to 20 poses.Source: C. Controlled Representation Inputs · pp. 10–11

Real-Vehicle Interface

  • The real-vehicle-derived evaluation reconstructs a Hualikan, Beijing delivery-vehicle parking site from an HD map and 1,000 operational clips containing 95,000 obstacle-detection frames.During evaluation, the planner replans once per recorded frame from target-relative ego, obstacle-polygon, and map-boundary geometry.Source: C. Controlled Representation Inputs · p. 11

Search Effort

  • Across 47 clips ranked by Hybrid A* search time, NeuralParker-Guided A* increasingly undercuts Hybrid A* as clip difficulty grows.Individual clips can exceed Hybrid A* time when the learned reference path disagrees with the search’s preferred maneuver.Source: C. Controlled Representation Inputs · p. 11

Related papers

This paper cites

Loading 2608.24485v1…