Selected work

Representative Papers

UniTeD: Unified Temporal Diffusion for Joint Perception and Planning in Autonomous Driving

Jun 24, 2026

This work addresses the challenges of error propagation, optimization difficulty, and insufficient robustness arising from the decoupling of perception and planning in autonomous driving. To this end, the authors propose UniTeD, a unified temporal diffusion framework that jointly models perception and planning within a shared generative space, enabling bidirectional information exchange through iterative denoising. Key innovations include a Temporal Transition Module (TTM) to mitigate noise-level mismatches, an Anchor Refresh Strategy (ARS) to align training and inference distributions, and a noise-conditioned multitask training mechanism. Evaluated on multiple autonomous driving benchmarks, UniTeD significantly outperforms existing discriminative end-to-end approaches and diffusion-based planning methods, achieving state-of-the-art performance.

0 citationsRead paper

VICR: Visual In-Context Restoration for Real-World Image Super-Resolution

May 30, 2026

Real-world image super-resolution faces the challenge of simultaneously preserving structural fidelity and generating realistic details, as existing methods often employ coupled conditional mechanisms that lead to structural distortions or semantic inconsistencies. This work reframes the task as an image inpainting problem and introduces a decoupled local–global visual prior injection mechanism within a diffusion Transformer (DiT) architecture, enabling separate capture of fine-grained and semantic-level cues from low-quality inputs to jointly restore structure and synthesize coherent details. Furthermore, an inference-time agent is incorporated to dynamically refine semantic prompts based on input visual evidence, without requiring model parameter updates. With only 127 million trainable parameters, the proposed method achieves state-of-the-art performance across multiple real-world super-resolution benchmarks.

0 citationsRead paper

SimPB++: Simultaneously Detecting 2D and 3D Objects from Multiple Cameras

May 03, 2026

This work addresses the challenge of unifying 2D perspective-view and 3D bird’s-eye-view (BEV) perception in multi-camera autonomous driving by proposing SimPB++, an end-to-end model featuring a hybrid decoder architecture. The method establishes a 3D–2D–3D iterative refinement loop through dynamic query assignment, adaptive query aggregation, and Query-group Attention. It further incorporates Crop-and-Scale and Propagating Denoising strategies to enable mixed supervision from both 2D-only and fully annotated data, substantially reducing reliance on costly 3D labels. Evaluated on nuScenes, SimPB++ achieves state-of-the-art performance in both 2D and 3D object detection, while demonstrating strong long-range detection capability—up to 150 meters—on Argoverse2.

0 citationsRead paper

TopoHR: Hierarchical Centerline Representation for Cyclic Topology Reasoning in Driving Scenes with Point-to-Instance Relations

Apr 27, 2026

Existing approaches to autonomous driving topology reasoning typically decouple centerline detection from topological inference and neglect point-to-instance (P2I) relationships, leading to inaccurate topological modeling. This work proposes TopoHR, an end-to-end hierarchical framework that unifies the modeling of P2I and inter-instance relationships through a hierarchical query mechanism involving point queries, instance queries, and semantic representations. By enabling iterative interactions between detection and reasoning modules, TopoHR achieves closed-loop co-optimization. Evaluated on the OpenLane-V2 benchmark, TopoHR demonstrates substantial improvements: on subset_A, it achieves a 3.8-point gain in DETₗ and a 5.4-point gain in TOPₗₗ; on subset_B, it improves DETₗ by 11.0 points and TOPₗₗ by 7.9 points.

0 citationsRead paper

DiffRefiner: Coarse to Fine Trajectory Planning via Diffusion Refinement with Semantic Interaction for End to End Autonomous Driving

Nov 21, 2025

To address the limited flexibility and accuracy in generative trajectory prediction caused by reliance on handcrafted anchors or random noise, this paper proposes a semantic-interaction-enhanced “coarse-to-fine” two-stage trajectory planning framework. In the first stage, a Transformer-based proposal module generates semantically aware initial trajectories; in the second stage, a fine-grained diffusion denoising decoder iteratively refines these proposals by fusing multimodal sensor inputs and environmental semantic cues. The method unifies discriminative proposal generation with generative refinement, significantly improving scene consistency and driving safety of predicted trajectories. Evaluated on NAVSIM v2 and Bench2Drive benchmarks, our approach achieves 87.4 EPDMS, 87.1 DS, and 71.4 SR—outperforming all prior methods and establishing new state-of-the-art results on both benchmarks.

0 citationsRead paper

Recent publications

Latest Papers

UniTeD: Unified Temporal Diffusion for Joint Perception and Planning in Autonomous Driving

Jun 24, 2026

This work addresses the challenges of error propagation, optimization difficulty, and insufficient robustness arising from the decoupling of perception and planning in autonomous driving. To this end, the authors propose UniTeD, a unified temporal diffusion framework that jointly models perception and planning within a shared generative space, enabling bidirectional information exchange through iterative denoising. Key innovations include a Temporal Transition Module (TTM) to mitigate noise-level mismatches, an Anchor Refresh Strategy (ARS) to align training and inference distributions, and a noise-conditioned multitask training mechanism. Evaluated on multiple autonomous driving benchmarks, UniTeD significantly outperforms existing discriminative end-to-end approaches and diffusion-based planning methods, achieving state-of-the-art performance.

0 citationsRead paper

VICR: Visual In-Context Restoration for Real-World Image Super-Resolution

May 30, 2026

Real-world image super-resolution faces the challenge of simultaneously preserving structural fidelity and generating realistic details, as existing methods often employ coupled conditional mechanisms that lead to structural distortions or semantic inconsistencies. This work reframes the task as an image inpainting problem and introduces a decoupled local–global visual prior injection mechanism within a diffusion Transformer (DiT) architecture, enabling separate capture of fine-grained and semantic-level cues from low-quality inputs to jointly restore structure and synthesize coherent details. Furthermore, an inference-time agent is incorporated to dynamically refine semantic prompts based on input visual evidence, without requiring model parameter updates. With only 127 million trainable parameters, the proposed method achieves state-of-the-art performance across multiple real-world super-resolution benchmarks.

0 citationsRead paper

SimPB++: Simultaneously Detecting 2D and 3D Objects from Multiple Cameras

May 03, 2026

This work addresses the challenge of unifying 2D perspective-view and 3D bird’s-eye-view (BEV) perception in multi-camera autonomous driving by proposing SimPB++, an end-to-end model featuring a hybrid decoder architecture. The method establishes a 3D–2D–3D iterative refinement loop through dynamic query assignment, adaptive query aggregation, and Query-group Attention. It further incorporates Crop-and-Scale and Propagating Denoising strategies to enable mixed supervision from both 2D-only and fully annotated data, substantially reducing reliance on costly 3D labels. Evaluated on nuScenes, SimPB++ achieves state-of-the-art performance in both 2D and 3D object detection, while demonstrating strong long-range detection capability—up to 150 meters—on Argoverse2.

0 citationsRead paper

TopoHR: Hierarchical Centerline Representation for Cyclic Topology Reasoning in Driving Scenes with Point-to-Instance Relations

Apr 27, 2026

Existing approaches to autonomous driving topology reasoning typically decouple centerline detection from topological inference and neglect point-to-instance (P2I) relationships, leading to inaccurate topological modeling. This work proposes TopoHR, an end-to-end hierarchical framework that unifies the modeling of P2I and inter-instance relationships through a hierarchical query mechanism involving point queries, instance queries, and semantic representations. By enabling iterative interactions between detection and reasoning modules, TopoHR achieves closed-loop co-optimization. Evaluated on the OpenLane-V2 benchmark, TopoHR demonstrates substantial improvements: on subset_A, it achieves a 3.8-point gain in DETₗ and a 5.4-point gain in TOPₗₗ; on subset_B, it improves DETₗ by 11.0 points and TOPₗₗ by 7.9 points.

0 citationsRead paper

DiffRefiner: Coarse to Fine Trajectory Planning via Diffusion Refinement with Semantic Interaction for End to End Autonomous Driving

Nov 21, 2025

To address the limited flexibility and accuracy in generative trajectory prediction caused by reliance on handcrafted anchors or random noise, this paper proposes a semantic-interaction-enhanced “coarse-to-fine” two-stage trajectory planning framework. In the first stage, a Transformer-based proposal module generates semantically aware initial trajectories; in the second stage, a fine-grained diffusion denoising decoder iteratively refines these proposals by fusing multimodal sensor inputs and environmental semantic cues. The method unifies discriminative proposal generation with generative refinement, significantly improving scene consistency and driving safety of predicted trajectories. Evaluated on NAVSIM v2 and Bench2Drive benchmarks, our approach achieves 87.4 EPDMS, 87.1 DS, and 71.4 SR—outperforming all prior methods and establishing new state-of-the-art results on both benchmarks.

0 citationsRead paper