Selected work

Representative Papers

DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation

Aug 12, 2026

This work addresses key limitations in aerial vision-and-language navigation—namely, restricted historical context, short planning horizons, and unreliable termination decisions—by introducing a novel approach that integrates a causal memory mechanism, receding-horizon diffusion-based planning, and a lightweight stop-detection module (LiteStop). The method enhances current visual representations with causally aligned historical memory to prevent future information leakage, employs a diffusion policy to predict K-step action sequences while executing only the first step to enable long-horizon planning, and directly estimates stopping probability from action logits. Built upon the Dream-VLA architecture, the proposed system achieves state-of-the-art performance on the OpenFly benchmark, attaining success rates of 32.04% and 29.46% in seen and unseen scenes, respectively, with corresponding SPL scores of 28.22% and 23.54%, and the lowest navigation error among existing methods.

0 citationsRead paper

DM3D: Deformable Mamba via Offset-Guided Gaussian Sequencing for Point Cloud Understanding

Dec 03, 2025

To resolve the fundamental conflict between point cloud disorder and the input-order dependency of State Space Models (SSMs), this paper proposes a deformable Mamba architecture. Methodologically, it introduces deformable scanning for point cloud serialization—first incorporating offset-guided Gaussian KNN resampling and differentiable Gaussian reordering to jointly achieve local geometric adaptivity and global structural awareness. A triple-path frequency fusion module is further designed to enhance spectral-domain modeling. The entire serialization process is end-to-end optimized, significantly improving SSMs’ capability to jointly capture long-range dependencies and fine-grained local structures in point clouds. Extensive experiments demonstrate state-of-the-art performance on classification, few-shot learning, and part segmentation tasks. Results validate that structure-adaptive serialization is pivotal for unlocking the full modeling potential of SSMs on point cloud data.

0 citationsRead paper

Hybrid-Emba3D: Geometry-Aware and Cross-Path Feature Hybrid Enhanced State Space Model for Point Cloud Classification

May 16, 2025

Point cloud classification faces dual challenges in balancing local geometric feature extraction with model complexity, while existing Mamba architectures—constrained by unidirectional state-space modeling—struggle to capture local spatial correlations inherent in unordered point clouds. To address this, we propose a Bidirectional Enhanced Mamba (BEM) architecture. Our method introduces geometric-aware local pooling and a geometry-feature coupling mechanism, along with a dual-path collaborative feature enhancer that overcomes the unidirectional limitation of standard state-space models (SSMs). We further incorporate parameter-free dynamic geometric information aggregation and cross-path feature mixing, significantly improving responsiveness to local geometric discontinuities and sparse structural signals. Evaluated on ModelNet40, BEM achieves 95.99% classification accuracy—setting a new state-of-the-art—while adding only 0.03M parameters.

0 citationsRead paper

Recent publications

Latest Papers

DreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation

Aug 12, 2026

This work addresses key limitations in aerial vision-and-language navigation—namely, restricted historical context, short planning horizons, and unreliable termination decisions—by introducing a novel approach that integrates a causal memory mechanism, receding-horizon diffusion-based planning, and a lightweight stop-detection module (LiteStop). The method enhances current visual representations with causally aligned historical memory to prevent future information leakage, employs a diffusion policy to predict K-step action sequences while executing only the first step to enable long-horizon planning, and directly estimates stopping probability from action logits. Built upon the Dream-VLA architecture, the proposed system achieves state-of-the-art performance on the OpenFly benchmark, attaining success rates of 32.04% and 29.46% in seen and unseen scenes, respectively, with corresponding SPL scores of 28.22% and 23.54%, and the lowest navigation error among existing methods.

0 citationsRead paper

DM3D: Deformable Mamba via Offset-Guided Gaussian Sequencing for Point Cloud Understanding

Dec 03, 2025

To resolve the fundamental conflict between point cloud disorder and the input-order dependency of State Space Models (SSMs), this paper proposes a deformable Mamba architecture. Methodologically, it introduces deformable scanning for point cloud serialization—first incorporating offset-guided Gaussian KNN resampling and differentiable Gaussian reordering to jointly achieve local geometric adaptivity and global structural awareness. A triple-path frequency fusion module is further designed to enhance spectral-domain modeling. The entire serialization process is end-to-end optimized, significantly improving SSMs’ capability to jointly capture long-range dependencies and fine-grained local structures in point clouds. Extensive experiments demonstrate state-of-the-art performance on classification, few-shot learning, and part segmentation tasks. Results validate that structure-adaptive serialization is pivotal for unlocking the full modeling potential of SSMs on point cloud data.

0 citationsRead paper

Hybrid-Emba3D: Geometry-Aware and Cross-Path Feature Hybrid Enhanced State Space Model for Point Cloud Classification

May 16, 2025

Point cloud classification faces dual challenges in balancing local geometric feature extraction with model complexity, while existing Mamba architectures—constrained by unidirectional state-space modeling—struggle to capture local spatial correlations inherent in unordered point clouds. To address this, we propose a Bidirectional Enhanced Mamba (BEM) architecture. Our method introduces geometric-aware local pooling and a geometry-feature coupling mechanism, along with a dual-path collaborative feature enhancer that overcomes the unidirectional limitation of standard state-space models (SSMs). We further incorporate parameter-free dynamic geometric information aggregation and cross-path feature mixing, significantly improving responsiveness to local geometric discontinuities and sparse structural signals. Evaluated on ModelNet40, BEM achieves 95.99% classification accuracy—setting a new state-of-the-art—while adding only 0.03M parameters.

0 citationsRead paper