Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Diffusion-Refined Planner with Reinforcement Learning Priors for Confined-Space Parking

Mingyang Jiang, Yueyuan Li, Jiaru Zhang, Songan Zhang, Ming Yang

TL;DR

This work tackles unreliable planning in confined-space parking by introducing DRIP, a diffusion-refined planner that leverages an RL-pretrained prior to regularize the diffusion process. The key idea is to align the diffusion training with a prior action distribution through a time-variant mean, enabling inference to start from a well-informed RL prior and undergo a short, targeted denoising refinement. The method couples an RL pretraining stage with a prior-aligned diffusion objective and a truncation strategy to ensure training–inference consistency, demonstrated to significantly boost success rates in narrow parking scenarios, especially under hard constraints. The results suggest practical impact for robust automated parking in tight urban environments and point to fruitful extensions combining diffusion refinement with online RL and policy improvement loops.

Abstract

The growing demand for parking has increased the need for automated parking planning methods that can operate reliably in confined spaces. In restricted and complex environments, high-precision maneuvers are required to achieve a high success rate in planning, yet existing approaches often rely on explicit action modeling, which faces challenges when accurately modeling the optimal action distribution. In this paper, we propose DRIP, a diffusion-refined planner anchored in reinforcement learning (RL) prior action distribution, in which an RL-pretrained policy provides prior action distributions to regularize the diffusion training process. During the inference phase the denoising process refines these coarse priors into more precise action distributions. By steering the denoising trajectory through the reinforcement learning prior distribution during training, the diffusion model inherits a well-informed initialization, resulting in more accurate action modeling, a higher planning success rate, and reduced inference steps. We evaluate our approach across parking scenarios with varying degrees of spatial constraints. Experimental results demonstrate that our method significantly improves planning performance in confined-space parking environments while maintaining strong generalization in common scenarios.

A Diffusion-Refined Planner with Reinforcement Learning Priors for Confined-Space Parking

TL;DR

This work tackles unreliable planning in confined-space parking by introducing DRIP, a diffusion-refined planner that leverages an RL-pretrained prior to regularize the diffusion process. The key idea is to align the diffusion training with a prior action distribution through a time-variant mean, enabling inference to start from a well-informed RL prior and undergo a short, targeted denoising refinement. The method couples an RL pretraining stage with a prior-aligned diffusion objective and a truncation strategy to ensure training–inference consistency, demonstrated to significantly boost success rates in narrow parking scenarios, especially under hard constraints. The results suggest practical impact for robust automated parking in tight urban environments and point to fruitful extensions combining diffusion refinement with online RL and policy improvement loops.

Abstract

The growing demand for parking has increased the need for automated parking planning methods that can operate reliably in confined spaces. In restricted and complex environments, high-precision maneuvers are required to achieve a high success rate in planning, yet existing approaches often rely on explicit action modeling, which faces challenges when accurately modeling the optimal action distribution. In this paper, we propose DRIP, a diffusion-refined planner anchored in reinforcement learning (RL) prior action distribution, in which an RL-pretrained policy provides prior action distributions to regularize the diffusion training process. During the inference phase the denoising process refines these coarse priors into more precise action distributions. By steering the denoising trajectory through the reinforcement learning prior distribution during training, the diffusion model inherits a well-informed initialization, resulting in more accurate action modeling, a higher planning success rate, and reduced inference steps. We evaluate our approach across parking scenarios with varying degrees of spatial constraints. Experimental results demonstrate that our method significantly improves planning performance in confined-space parking environments while maintaining strong generalization in common scenarios.
Paper Structure (19 sections, 27 equations, 5 figures, 3 tables, 2 algorithms)

This paper contains 19 sections, 27 equations, 5 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Work
  3. Parking Path-planning
  4. Diffusion Policy
  5. Preliminaries
  6. Problem Formulation
  7. Diffusion Policy Preliminaries
  8. Methodology
  9. Diffusion-refinement Framework
  10. Prior-Aligned Diffusion
  11. Introduction of Time-Variant Mean
  12. Training Objective
  13. Selection of Coefficients
  14. Training–Inference Pipeline and Learning Objectives
  15. Experiment
  16. ...and 4 more sections

Figures (5)

  • Figure 1: Comparison between explicit action modeling and diffusion-based policy for parking. (a) An explicit action model parameterizes an action distribution or probabilistic function and selects actions by sampling or maximizing this distribution. This can leave nontrivial mass on suboptimal maneuvers, which in confined spaces leads to collisions or additional corrective steps (b). (c) A diffusion policy models actions implicitly via learning a gradient field and iteratively refines noisy proposals toward feasible, high-precision trajectories.
  • Figure 2: Overview of the proposed framework: (1) pretrain an RL policy to obtain a state-conditional prior distribution; (2) roll out this policy to collect a dataset; (3) train a diffusion policy on the collected dataset to refine the prior (via IL or RL).
  • Figure 3: Illustration for the proposed prior-aligned diffusion training. Standard diffusion starts from random noise and yields an intermediate distribution that does not align with the RL-pretrained prior distribution. Our goal is to match the distribution from the denoising process with the prior distribution through alignment training, so the denoising process can start from the prior distribution at inference.
  • Figure 4: Visualization for a vertical parking scenario. (a) Scene overview. (b) A reasonable path planning for one-time park-in. (c) Feasibility of 400 uniformly sampled actions in the local action space; actions enabling a single-maneuver park-in with one gear change are highlighted in green/yellow. Others are in red. (d) RL-pretrained action distribution at the same position. (e) Random action initialization from diffusion policy before denoising. (f) Action distribution with diffusion policy after denoising. (g) Proposed diffusion refinement initialized from the RL-pretrained prior distribution; (h) Final action distribution after prior-aligned denoising.
  • Figure 5: Comparison of average planning success rate (%) and inference time (ms) between different methods.