Selected work

Representative Papers

StrokeNet: Unveiling How to Learn Fine-Grained Interactions in Online Handwritten Stroke Classification

Dec 05, 2025

Online handwritten stroke classification faces challenges in modeling fine-grained semantic relationships due to high variability in writing styles, ambiguous content, and dynamic spatial positioning. To address the limitation of existing methods in capturing local stroke interactions, this paper proposes a reference-point sequence coupling representation and an Inline Sequence Attention module, incorporating a Cross-Ellipse Query mechanism for multi-scale spatial feature aggregation. A joint optimization framework is introduced to simultaneously predict stroke categories and semantic transition relations. The method integrates dynamic reference-point selection, sequential modeling, spatial query clustering, and multi-task learning (primary classification + regression + auxiliary branch), enabling end-to-end training. Evaluated on public benchmarks including CASIA-onDo, it achieves state-of-the-art performance—improving accuracy from 93.81% to 95.54%—with significantly enhanced robustness and generalization capability.

0 citationsRead paper

SAM2Grasp: Resolve Multi-modal Grasping via Prompt-conditioned Temporal Action Prediction

Dec 02, 2025

Imitation learning for robotic grasping is often plagued by the multimodal problem: when a scene contains multiple valid targets, demonstrations of grasping different objects create conflicting training signals. Standard imitation learning policies fail by averaging these distinct actions into a single, invalid action. In this paper, we introduce SAM2Grasp, a novel framework that resolves this issue by reformulating the task as a uni-modal, prompt-conditioned prediction problem. Our method leverages the frozen SAM2 model to use its powerful visual temporal tracking capability and introduces a lightweight, trainable action head that operates in parallel with its native segmentation head. This design allows for training only the small action head on pre-computed temporal-visual features from SAM2. During inference, an initial prompt, such as a bounding box provided by an upstream object detection model, designates the specific object to be grasped. This prompt conditions the action head to predict a unique, unambiguous grasp trajectory for that object alone. In all subsequent video frames, SAM2's built-in temporal tracking capability automatically maintains stable tracking of the selected object, enabling our model to continuously predict the grasp trajectory from the video stream without further external guidance. This temporal-prompted approach effectively eliminates ambiguity from the visuomotor policy. We demonstrate through extensive experiments that SAM2Grasp achieves state-of-the-art performance in cluttered, multi-object grasping tasks.

0 citationsRead paper

Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis

Sep 11, 2025

Existing large vision-language models (LVLMs) perform well on general medical tasks but exhibit significant limitations in dental panoramic radiograph analysis due to the lack of domain-specific multimodal data and standardized evaluation benchmarks tailored to dense anatomical structures and subtle pathological features. To address this gap, we introduce MMOral—the first multimodal instruction-tuning dataset for oral imaging—comprising 20,563 panoramic X-ray images and 1.3 million expert-curated instructions—alongside MMOral-Bench, a comprehensive evaluation benchmark. We further propose OralGPT, a lightweight fine-tuned variant of Qwen2.5-VL-7B, achieving a 24.73% improvement over the baseline after single-round supervised fine-tuning. Notably, GPT-4o scores only 41.45% on MMOral-Bench, underscoring the task’s difficulty and the necessity of domain adaptation. This work establishes the first reproducible foundation for dental multimodal AI, providing a benchmarked dataset, rigorous evaluation protocol, and effective methodology for specialized medical LVM development.

0 citationsRead paper

Train Once, Deploy Anywhere: Realize Data-Efficient Dynamic Object Manipulation

Aug 19, 2025

To address the high cost of demonstration data collection and poor generalization in dynamic object manipulation, this paper proposes the Generalizable Entropy-based Manipulation (GEM) framework—the first to systematically integrate entropy theory into imitation learning, establishing an information-theoretic generalization optimization mechanism. GEM jointly incorporates entropy-driven behavioral representation, invariance modeling across morphology, dynamics, and geometry, and a sim-to-real transfer strategy, substantially reducing reliance on scene-specific demonstrations. Without on-site fine-tuning, GEM successfully executed over 10,000 operations in a real-world cafeteria utensil retrieval task, achieving ≥97% success rate—demonstrating strong cross-domain adaptability and deployment robustness in complex dynamic environments. The core contribution is the establishment of the first entropy-theoretic framework explicitly designed for generalization optimization, enabling high-reliability dynamic manipulation with zero on-site demonstrations.

0 citationsRead paper

Bootstrapping Imitation Learning for Long-horizon Manipulation via Hierarchical Data Collection Space

May 23, 2025

To address the challenges of imitation learning in long-horizon robotic manipulation—namely, its heavy reliance on large-scale human demonstrations and poor generalization—this paper proposes the Hierarchical Data Collection Space (HD-Space) framework. HD-Space decomposes complex tasks into atomic subtasks and constructs a structured state-action space to enable high-quality, low-overhead demonstration generation. Crucially, it enhances demonstration robustness at the data collection source, enabling end-to-end policy training from only a small number of high-information demonstrations—thereby substantially reducing dependence on data volume and human intervention. Evaluated on two simulated and five real-world long-horizon manipulation tasks, HD-Space consistently outperforms mainstream baselines in success rate, demonstrating simultaneous improvements in data efficiency and policy performance.

0 citationsRead paper

Recent publications

Latest Papers

StrokeNet: Unveiling How to Learn Fine-Grained Interactions in Online Handwritten Stroke Classification

Dec 05, 2025

Online handwritten stroke classification faces challenges in modeling fine-grained semantic relationships due to high variability in writing styles, ambiguous content, and dynamic spatial positioning. To address the limitation of existing methods in capturing local stroke interactions, this paper proposes a reference-point sequence coupling representation and an Inline Sequence Attention module, incorporating a Cross-Ellipse Query mechanism for multi-scale spatial feature aggregation. A joint optimization framework is introduced to simultaneously predict stroke categories and semantic transition relations. The method integrates dynamic reference-point selection, sequential modeling, spatial query clustering, and multi-task learning (primary classification + regression + auxiliary branch), enabling end-to-end training. Evaluated on public benchmarks including CASIA-onDo, it achieves state-of-the-art performance—improving accuracy from 93.81% to 95.54%—with significantly enhanced robustness and generalization capability.

0 citationsRead paper

SAM2Grasp: Resolve Multi-modal Grasping via Prompt-conditioned Temporal Action Prediction

Dec 02, 2025

Imitation learning for robotic grasping is often plagued by the multimodal problem: when a scene contains multiple valid targets, demonstrations of grasping different objects create conflicting training signals. Standard imitation learning policies fail by averaging these distinct actions into a single, invalid action. In this paper, we introduce SAM2Grasp, a novel framework that resolves this issue by reformulating the task as a uni-modal, prompt-conditioned prediction problem. Our method leverages the frozen SAM2 model to use its powerful visual temporal tracking capability and introduces a lightweight, trainable action head that operates in parallel with its native segmentation head. This design allows for training only the small action head on pre-computed temporal-visual features from SAM2. During inference, an initial prompt, such as a bounding box provided by an upstream object detection model, designates the specific object to be grasped. This prompt conditions the action head to predict a unique, unambiguous grasp trajectory for that object alone. In all subsequent video frames, SAM2's built-in temporal tracking capability automatically maintains stable tracking of the selected object, enabling our model to continuously predict the grasp trajectory from the video stream without further external guidance. This temporal-prompted approach effectively eliminates ambiguity from the visuomotor policy. We demonstrate through extensive experiments that SAM2Grasp achieves state-of-the-art performance in cluttered, multi-object grasping tasks.

0 citationsRead paper

Towards Better Dental AI: A Multimodal Benchmark and Instruction Dataset for Panoramic X-ray Analysis

Sep 11, 2025

Existing large vision-language models (LVLMs) perform well on general medical tasks but exhibit significant limitations in dental panoramic radiograph analysis due to the lack of domain-specific multimodal data and standardized evaluation benchmarks tailored to dense anatomical structures and subtle pathological features. To address this gap, we introduce MMOral—the first multimodal instruction-tuning dataset for oral imaging—comprising 20,563 panoramic X-ray images and 1.3 million expert-curated instructions—alongside MMOral-Bench, a comprehensive evaluation benchmark. We further propose OralGPT, a lightweight fine-tuned variant of Qwen2.5-VL-7B, achieving a 24.73% improvement over the baseline after single-round supervised fine-tuning. Notably, GPT-4o scores only 41.45% on MMOral-Bench, underscoring the task’s difficulty and the necessity of domain adaptation. This work establishes the first reproducible foundation for dental multimodal AI, providing a benchmarked dataset, rigorous evaluation protocol, and effective methodology for specialized medical LVM development.

0 citationsRead paper

Train Once, Deploy Anywhere: Realize Data-Efficient Dynamic Object Manipulation

Aug 19, 2025

To address the high cost of demonstration data collection and poor generalization in dynamic object manipulation, this paper proposes the Generalizable Entropy-based Manipulation (GEM) framework—the first to systematically integrate entropy theory into imitation learning, establishing an information-theoretic generalization optimization mechanism. GEM jointly incorporates entropy-driven behavioral representation, invariance modeling across morphology, dynamics, and geometry, and a sim-to-real transfer strategy, substantially reducing reliance on scene-specific demonstrations. Without on-site fine-tuning, GEM successfully executed over 10,000 operations in a real-world cafeteria utensil retrieval task, achieving ≥97% success rate—demonstrating strong cross-domain adaptability and deployment robustness in complex dynamic environments. The core contribution is the establishment of the first entropy-theoretic framework explicitly designed for generalization optimization, enabling high-reliability dynamic manipulation with zero on-site demonstrations.

0 citationsRead paper

Bootstrapping Imitation Learning for Long-horizon Manipulation via Hierarchical Data Collection Space

May 23, 2025

To address the challenges of imitation learning in long-horizon robotic manipulation—namely, its heavy reliance on large-scale human demonstrations and poor generalization—this paper proposes the Hierarchical Data Collection Space (HD-Space) framework. HD-Space decomposes complex tasks into atomic subtasks and constructs a structured state-action space to enable high-quality, low-overhead demonstration generation. Crucially, it enhances demonstration robustness at the data collection source, enabling end-to-end policy training from only a small number of high-information demonstrations—thereby substantially reducing dependence on data volume and human intervention. Evaluated on two simulated and five real-world long-horizon manipulation tasks, HD-Space consistently outperforms mainstream baselines in success rate, demonstrating simultaneous improvements in data efficiency and policy performance.

0 citationsRead paper