pith. sign in

arxiv: 2605.22722 · v1 · pith:QUFGWFG5new · submitted 2026-05-21 · 💻 cs.RO · cs.SY· eess.SY

N3P: Accelerated Automated Parking via a Learning-Based Naturalistic Three-Stage Scheme

Yifan Xue , Toktam Mohammadnejad , Faizan M Tariq , Sangjae Bae , David Isele , Yosuke Sakamoto , Nadia Figueroa , Jovin D'sa This is my paper

Pith reviewed 2026-05-22 04:58 UTC · model grok-4.3

classification 💻 cs.RO cs.SYeess.SY
keywords automated parkingpath planningHybrid A*learning-based planningthree-stage schemepreparatory posereinforcement learning baselinekinematic feasibility
0
0 comments X

The pith

A learned preparatory pose splits complex parking maneuvers so Hybrid A* can plan them over 80 percent faster.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents N3P, a three-stage planning method for autonomous parking that inserts a machine-learned intermediate vehicle pose between the starting position and the final parked spot. This decomposition turns one long, geometrically difficult search into two shorter searches that Hybrid A* can solve much more quickly while still guaranteeing kinematic feasibility and collision-free motion. Experiments in perpendicular and parallel parking show the enhanced planner runs more than 80 percent faster than plain Hybrid A*, produces shorter trajectories with fewer gear changes, and outperforms reinforcement-learning baselines on success rate and path quality. The approach keeps the reliability of classical search-based planning but uses the learned shortcut to reach real-time speeds in tight spaces. If the predicted preparatory pose is accurate, the method makes high-quality, safe parking feasible on current vehicle hardware.

Core claim

N3P decomposes an automated parking maneuver into three stages by using a learning module to predict a preparatory pose; the two resulting simpler subproblems can then be solved independently by Hybrid A*, yielding feasible, collision-free trajectories with substantially lower computational cost than a single-stage search.

What carries the argument

A learning module that predicts an intermediate preparatory pose to split one complex parking maneuver into two simpler subproblems for Hybrid A*.

If this is right

  • N3P-enhanced Hybrid A* reduces planning time by more than 80 percent in both perpendicular and parallel parking.
  • The resulting trajectories are shorter and contain fewer gear changes than those generated by reinforcement-learning baselines.
  • Success rate and overall trajectory quality remain higher than RL methods while planning time stays comparable or lower.
  • Kinematic feasibility and collision avoidance are preserved because the underlying Hybrid A* solver is unchanged.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same preparatory-pose idea could be tested on other constrained maneuvers such as garage docking or narrow-passage navigation.
  • If the predictor generalizes across vehicle sizes and turning radii, it might reduce the need for per-vehicle retuning of the planner.
  • Real-world use would still require handling moving obstacles and sensor uncertainty that the current static-environment tests do not address.

Load-bearing premise

The learning module must accurately predict a preparatory pose that genuinely simplifies the overall maneuver without introducing new collision risks or requiring extensive retraining for different environments.

What would settle it

A test set of parking scenarios in which the predicted preparatory pose repeatedly produces paths that either collide with obstacles or require more total planning time than the single-stage baseline would falsify the central speedup claim.

Figures

Figures reproduced from arXiv: 2605.22722 by David Isele, Faizan M Tariq, Jovin D'sa, Nadia Figueroa, Sangjae Bae, Toktam Mohammadnejad, Yifan Xue, Yosuke Sakamoto.

Figure 1
Figure 1. Figure 1: During parking, human drivers often follow a prepa [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Examples of parking-spot-based frames pˆ for forward, reverse, and parallel parking. Drive lanes are shown in green and parking spots in yellow. Here l is the wheelbase of the vehicle, approximated as a rectangle. u = [v, δ] ⊤ is the control input, where v is the longitudinal velocity and δ is the steering angle of the front wheel. The control inputs are subject to box constraints: v ∈ [vmin, vmax] and δ ∈… view at source ↗
Figure 3
Figure 3. Figure 3: Left: extraction of abstracted parameters from an arbi [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Reeds–Shepp start poses for the reverse parking [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Comparison of KNN-N3P (top) and the RL agent trained on the evaluation environment (bottom) in forward, parallel, [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
read the original abstract

Autonomous parking requires efficient path planning that ensures kinematic feasibility and collision avoidance in constrained environments. Hybrid A* is widely used but computationally expensive, while reinforcement learning (RL) methods lack reliability and often struggle with long-horizon geometric constraints, leading to suboptimal trajectories. We present N3P, a fast learning-based three-stage framework for automated parking. By introducing an intermediate preparatory pose and using a learning module to predict it, N3P decomposes the maneuver into simpler subproblems, thereby reducing computational complexity and accelerating path generation. We validate the framework by integrating it with Hybrid A* algorithms. Experiments in perpendicular and parallel parking scenarios show that N3P-enhanced Hybrid A* speeds up planning by more than 80%. It also outperforms RL baselines in success rate and trajectory quality, producing shorter trajectories with fewer gear changes, while achieving comparable or lower planning time in most cases.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The manuscript proposes N3P, a learning-based naturalistic three-stage scheme for automated parking. A learning module predicts an intermediate preparatory pose that decomposes the maneuver into simpler subproblems, which is then integrated with Hybrid A* to accelerate path planning while ensuring kinematic feasibility and collision avoidance. Experiments in perpendicular and parallel parking report >80% planning speedup over standard Hybrid A*, plus higher success rates, shorter trajectories, and fewer gear changes than RL baselines.

Significance. If the performance claims hold under rigorous validation, the hybrid decomposition approach could meaningfully advance real-time autonomous parking by mitigating the computational cost of search-based planners and the long-horizon reliability issues of pure RL. The explicit integration with Hybrid A* and direct comparison to RL baselines on trajectory quality metrics are constructive elements.

major comments (2)
  1. Abstract: the central claim that N3P-enhanced Hybrid A* 'speeds up planning by more than 80%' is presented without any description of the learning module's network architecture, training procedure, dataset, or statistical tests. This absence prevents verification of the quantitative gains and reproducibility of the reported advantage.
  2. Method (three-stage pipeline): no explicit feasibility or collision check on the predicted preparatory pose is described. If the pose lies in an unreachable or colliding region, the three-stage scheme either fails or falls back to full-horizon search, directly undermining the speedup and trajectory-quality claims.
minor comments (1)
  1. Abstract: the phrase 'naturalistic' is used in the title but never defined or operationalized; a one-sentence clarification would improve precision.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify key aspects of the presentation. We address each major comment below and indicate the corresponding revisions.

read point-by-point responses

  1. Referee: Abstract: the central claim that N3P-enhanced Hybrid A* 'speeds up planning by more than 80%' is presented without any description of the learning module's network architecture, training procedure, dataset, or statistical tests. This absence prevents verification of the quantitative gains and reproducibility of the reported advantage.

    Authors: The abstract is intentionally concise and summarizes the main results. Full details on the network architecture (a convolutional encoder followed by fully connected layers), training procedure (supervised learning on expert trajectories), dataset (simulated parking scenarios with 50k samples), and statistical tests (paired t-tests on planning time across 500 trials) are provided in Sections 3.2 and 4.2. To address the concern directly, we will expand the abstract with a one-sentence reference to the learning module and add explicit pointers to these sections. revision: yes

  2. Referee: Method (three-stage pipeline): no explicit feasibility or collision check on the predicted preparatory pose is described. If the pose lies in an unreachable or colliding region, the three-stage scheme either fails or falls back to full-horizon search, directly undermining the speedup and trajectory-quality claims.

    Authors: We agree that an explicit check strengthens the method. The learning module is trained exclusively on collision-free and kinematically feasible poses extracted from Hybrid A* solutions, which provides implicit safety within the training distribution. However, to handle out-of-distribution predictions, we will add a post-prediction verification step that reuses the same collision checker and kinematic constraints as the downstream Hybrid A* planner; invalid poses trigger an immediate fallback to standard Hybrid A*. This will be described in the revised Section 3.3. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical validation stands independently of any self-referential definitions.

full rationale

The N3P framework is presented as a three-stage empirical pipeline that trains a learning module to predict an intermediate preparatory pose, then invokes Hybrid A* on the resulting subproblems. No equations, uniqueness theorems, or fitted parameters are shown to reduce by construction to the authors' own prior outputs or self-citations. The speedup and performance claims rest on experimental comparisons against standard baselines rather than on any definitional equivalence or load-bearing self-reference. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review provides no explicit free parameters, axioms, or invented entities. The learning module presumably contains trained weights, but none are named or quantified here.

pith-pipeline@v0.9.0 · 5717 in / 1143 out tokens · 39506 ms · 2026-05-22T04:58:29.746997+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages

  1. [1]

    Improved hierarchical a-star algorithm for optimal parking path planning of the large parking lot

    Liping Cheng, Chuanxi Liu, and Bo Yan. Improved hierarchical a-star algorithm for optimal parking path planning of the large parking lot. In2014 IEEE international conference on information and automation (ICIA), pages 695–698. IEEE, 2014

    work page 2014

  2. [2]

    Unified path planner for parking an autonomous vehicle based on rrt

    Long Han, Quoc Huy Do, and Seiichi Mita. Unified path planner for parking an autonomous vehicle based on rrt. In2011 IEEE International Conference on Robotics and Automation, pages 5622–

    work page

  3. [3]

    Hart, Nils J

    Peter E. Hart, Nils J. Nilsson, and Bertram Raphael. A formal basis for the heuristic determination of minimum cost paths.IEEE Transactions on Systems Science and Cybernetics, 4(2):100–107, 1968

    work page 1968

  4. [4]

    Rapidly-exploring random trees: A new tool for path planning.Research Report 9811, 1998

    Steven LaValle. Rapidly-exploring random trees: A new tool for path planning.Research Report 9811, 1998

    work page 1998

  5. [5]

    Practical search techniques in path planning for autonomous driving.ann arbor, 1001(48105):18–80, 2008

    Dmitri Dolgov, Sebastian Thrun, Michael Montemerlo, and James Diebel. Practical search techniques in path planning for autonomous driving.ann arbor, 1001(48105):18–80, 2008

    work page 2008

  6. [6]

    Shortest paths for the reeds- shepp car: a worked out example of the use of geometric techniques in nonlinear optimal control, 1991

    H ´ector J Sussmann and Guoqing Tang. Shortest paths for the reeds- shepp car: a worked out example of the use of geometric techniques in nonlinear optimal control, 1991

    work page 1991

  7. [7]

    Accelerated reeds–shepp and underspecified reeds–shepp algorithms for mobile robot path planning.IEEE Transactions on Robotics, 41:2691–2708, 2025

    Ibrahim Ibrahim, Wilm Decr ´e, and Jan Swevers. Accelerated reeds–shepp and underspecified reeds–shepp algorithms for mobile robot path planning.IEEE Transactions on Robotics, 41:2691–2708, 2025

    work page 2025

  8. [8]

    Optimization-based motion planning for autonomous parking considering dynamic obstacle: A hierarchical framework

    Xuemin Chi, Zhitao Liu, Jihao Huang, Feng Hong, and Hongye Su. Optimization-based motion planning for autonomous parking considering dynamic obstacle: A hierarchical framework. In2022 34th Chinese Control and Decision Conference (CCDC), pages 6229–6234, 2022

    work page 2022

  9. [9]

    Search-based path planning algorithm for autonomous parking: Multi-heuristic hybrid a

    Jihao Huang, Zhitao Liu, Xuemin Chi, Feng Hong, and Hongye Su. Search-based path planning algorithm for autonomous parking: Multi-heuristic hybrid a. In2022 34th Chinese Control and Decision Conference (CCDC), pages 6248–6253. IEEE, 2022

    work page 2022

  10. [10]

    Informed hybrid a star-based path planning algorithm in unstructured environments

    Antonio Acernese, Giulio Borrello, Luca Lorusso, and Michele Basso. Informed hybrid a star-based path planning algorithm in unstructured environments. In2024 European Control Conference (ECC), pages 1039–1044, 2024

    work page 2024

  11. [11]

    Optimization-based collision avoidance.IEEE Transactions on Control Systems Technology, 29(3):972–983, 2020

    Xiaojing Zhang, Alexander Liniger, and Francesco Borrelli. Optimization-based collision avoidance.IEEE Transactions on Control Systems Technology, 29(3):972–983, 2020

    work page 2020

  12. [12]

    Rehg, and Evangelos A

    Grady Williams, Paul Drews, Brian Goldfain, James M. Rehg, and Evangelos A. Theodorou. Aggressive driving with model predictive path integral control. In2016 IEEE International Conference on Robotics and Automation (ICRA), pages 1433–1440, 2016

    work page 2016

  13. [13]

    Eric Tseng

    Chang Liu, Seungho Lee, Scott Varnhagen, and H. Eric Tseng. Path planning for autonomous vehicles using model predictive control. In 2017 IEEE Intelligent Vehicles Symposium (IV), pages 174–179, 2017

    work page 2017

  14. [14]

    Automatic parking trajectory planning in narrow spaces based on hybrid a* and nmpc.Scientific Reports, 15(1):1384, 2025

    Pei Zhang, SiLong Zhou, Jie Hu, WenLong Zhao, Jiachen Zheng, Zhiling Zhang, and Chongzhi Gao. Automatic parking trajectory planning in narrow spaces based on hybrid a* and nmpc.Scientific Reports, 15(1):1384, 2025

    work page 2025

  15. [15]

    A rapid iterative trajectory planning method for automated parking through differential flatness.Robotics and Autonomous Systems, 182:104816, 2024

    Zhouheng Li, Lei Xie, Cheng Hu, and Hongye Su. A rapid iterative trajectory planning method for automated parking through differential flatness.Robotics and Autonomous Systems, 182:104816, 2024

    work page 2024

  16. [16]

    An efficient spatial-temporal trajectory planner for autonomous vehicles in unstructured environments.IEEE Transactions on Intelligent Transportation Systems, 25(2):1797–1814, 2024

    Zhichao Han, Yuwei Wu, Tong Li, Lu Zhang, Liuao Pei, Long Xu, Chengyang Li, Changjia Ma, Chao Xu, Shaojie Shen, and Fei Gao. An efficient spatial-temporal trajectory planner for autonomous vehicles in unstructured environments.IEEE Transactions on Intelligent Transportation Systems, 25(2):1797–1814, 2024

    work page 2024

  17. [17]

    Bai Li, Tankut Acarman, Youmin Zhang, Yakun Ouyang, Cagdas Ya- man, Qi Kong, Xiang Zhong, and Xiaoyan Peng. Optimization-based trajectory planning for autonomous parking with irregularly placed obstacles: A lightweight iterative framework.IEEE Transactions on Intelligent Transportation Systems, 23(8):11970–11981, 2022

    work page 2022

  18. [18]

    Runqi Chai, Antonios Tsourdos, Al Savvaris, Senchun Chai, Yuanqing Xia, and CL Philip Chen. Design and implementation of deep neural network-based control for automatic parking maneuver process.IEEE Transactions on Neural Networks and Learning Systems, 33(4):1400– 1413, 2020

    work page 2020

  19. [19]

    Deep learning-based trajectory planning and control for autonomous ground vehicle parking maneuver.IEEE Transactions on Automation Science and Engineering, 20(3):1633– 1647, 2022

    Runqi Chai, Derong Liu, Tianhao Liu, Antonios Tsourdos, Yuanqing Xia, and Senchun Chai. Deep learning-based trajectory planning and control for autonomous ground vehicle parking maneuver.IEEE Transactions on Automation Science and Engineering, 20(3):1633– 1647, 2022

    work page 2022

  20. [20]

    Neural motion planning for autonomous parking.International Journal of Control, Automation and Systems, 21(4):1309–1318, 2023

    Dongchan Kim and Kunsoo Huh. Neural motion planning for autonomous parking.International Journal of Control, Automation and Systems, 21(4):1309–1318, 2023

    work page 2023

  21. [21]

    Multi-vehicle conflict resolution in highly constrained spaces by merging optimal control and reinforce- ment learning.IFAC-PapersOnLine, 56(2):3308–3313, 2023

    Xu Shen and Francesco Borrelli. Multi-vehicle conflict resolution in highly constrained spaces by merging optimal control and reinforce- ment learning.IFAC-PapersOnLine, 56(2):3308–3313, 2023

    work page 2023

  22. [22]

    Runqi Chai, Hanlin Niu, Joaquin Carrasco, Farshad Arvin, Hujun Yin, and Barry Lennox. Design and experimental validation of deep reinforcement learning-based fast trajectory planning and control for mobile robot in unknown environment.IEEE Transactions on Neural Networks and Learning Systems, 35(4):5778–5792, 2022

    work page 2022

  23. [23]

    Hope: A reinforcement learning-based hybrid policy path planner for diverse parking scenarios.IEEE Transactions on Intelligent Transportation Systems, 26(5):6130–6141, 2025

    Mingyang Jiang, Yueyuan Li, Songan Zhang, Siyuan Chen, Chunxiang Wang, and Ming Yang. Hope: A reinforcement learning-based hybrid policy path planner for diverse parking scenarios.IEEE Transactions on Intelligent Transportation Systems, 26(5):6130–6141, 2025

    work page 2025

  24. [24]

    Mpc-based motion planning for autonomous truck-trailer maneuver- ing.IFAC-PapersOnLine, 56(2):4877–4882, 2023

    Mathias Bos, Bastiaan Vandewal, Wilm Decr ´e, and Jan Swevers. Mpc-based motion planning for autonomous truck-trailer maneuver- ing.IFAC-PapersOnLine, 56(2):4877–4882, 2023. 22nd IFAC World Congress

    work page 2023

  25. [25]

    Hierarchical parking path planning based on optimal parking positions

    Yaogang Zhang, Guoying Chen, Hongyu Hu, and Zhenhai Gao. Hierarchical parking path planning based on optimal parking positions. Automotive Innovation, 6(2):220–230, 2023

    work page 2023

  26. [26]

    Springer, 2006

    Rajesh Rajamani.Vehicle dynamics and control. Springer, 2006

    work page 2006

  27. [27]

    Guided hybrid a-star path planning algorithm for valet parking applications

    Saeid Sedighi, Duong-Van Nguyen, and Klaus-Dieter Kuhnert. Guided hybrid a-star path planning algorithm for valet parking applications. In2019 5th International Conference on Control, Automation and Robotics (ICCAR), pages 570–575, 2019

    work page 2019

  28. [28]

    Optimal paths for a car that goes both forwards and backwards.Pacific journal of mathematics, 145(2):367–393, 1990

    James Reeds and Lawrence Shepp. Optimal paths for a car that goes both forwards and backwards.Pacific journal of mathematics, 145(2):367–393, 1990

    work page 1990

  29. [29]

    Accessed: Jan

    Los angeles municipal code §12.21(a)(1): Parking stall dimen- sions.https://codelibrary.amlegal.com/codes/los_ angeles/latest/lapz/0-0-0-5183. Accessed: Jan. 2026

    work page 2026

  30. [30]

    2024 honda accord: Specifications & features (press release), 2024

    American Honda Motor Co., Inc. 2024 honda accord: Specifications & features (press release), 2024. Accessed: Oct. 31, 2025

    work page 2024