Dino-Diffusion Modular Designs Bridge the Cross-Domain Gap in Autonomous Parking

TL;DR

This work targets the cross-domain robustness problem in autonomous parking by decoupling perception, planning, and control. It introduces Dino-Diffusion Parking (DDP), a modular pipeline that combines a Visual Foundation Model for generalized perception, hindsight relabeling for robust target fusion, and a diffusion model operating in $SE(2)$ space for consistent trajectory planning, followed by Stanley control for precise tracking. Evaluated in CARLA across weather shifts and in a 3D Gaussian Splatting (3DGS) environment, DDP achieves parking success well above 90% in OOD scenarios and outperforms end-to-end baselines, with ablations validating the contributions of VFMs, relabeling, and diffusion planning. The results demonstrate strong sim-to-real transfer potential and point to practical impact in robust, data-efficient autonomous parking in diverse real-world conditions.

Abstract

Parking is a critical pillar of driving safety. While recent end-to-end (E2E) approaches have achieved promising in-domain results, robustness under domain shifts (e.g., weather and lighting changes) remains a key challenge. Rather than relying on additional data, in this paper, we propose Dino-Diffusion Parking (DDP), a domain-agnostic autonomous parking pipeline that integrates visual foundation models with diffusion-based planning to enable generalized perception and robust motion planning under distribution shifts. We train our pipeline in CARLA at regular setting and transfer it to more adversarial settings in a zero-shot fashion. Our model consistently achieves a parking success rate above 90% across all tested out-of-distribution (OOD) scenarios, with ablation studies confirming that both the network architecture and algorithmic design significantly enhance cross-domain performance over existing baselines. Furthermore, testing in a 3D Gaussian splatting (3DGS) environment reconstructed from a real-world parking lot demonstrates promising sim-to-real transfer.

Figures (6)

• Figure 1: Zero-Shot Domain Generalization: Our method addresses the challenge of training a parking policy under default weather conditions (e.g., CARLA sunny-noon, left) and deploying it zero-shot in other weather or illumination settings (middle) or in a 3DGS environment reconstructed from the real world (right). Our method (blue trajectories indicated by 'T') overcomes the cross-domain failures (red trajectories) observed in prior in-domain E2E pipelines and demonstrates promising zero-shot transfer to 3DGS testing.
• Figure 2: Dino-Diffusion Parking (DDP) Pipeline: We use DINOv2 to generate robust visual features and transform features into vehicle BEV. Then we relabel the target parking spot (in training only) and fuse with BEV. Finally, the diffusion planner employs a U-Net conditioned on the target pose and the fused feature to predict the future trajectory in $\mathrm{SE(2)}$ and execute it with Stanley control. The supervision signal is from ground truth depth, segmentation and expert trajectories, and it should be noted that the perception and planning are decoupled where the gradient doesn't pass from diffusion to fused BEV.
• Figure 3: Tracking with Stanley Control: Given the trajectory from the diffusion planner, the controller minimizes the heading error and distance error with a reference point.
• Figure 4: Robust Features from Vision Foundation Models: We utilize UMap 2018arXivUMAP to visualize the features at two stages, directly from the backbone and before they are projected into BEV. Compared to the features from the baseline model, our features from DINOv2 oquab2023dinov2 backbone are more consistent across different domains, highlighting the potential to reduce domain gap with foundation models in perception.
• Figure 5: