Selected work

Representative Papers

Self Supervised Learning from Automatically Generated Demonstrations for Visual Robotic Manipulation

Aug 01, 2026

This work proposes a self-supervised visual manipulation method that eliminates the need for manual programming, human demonstrations, or extrinsic camera-robot calibration, which traditionally incur high deployment costs. The robot autonomously generates visual demonstrations near the target pose and learns relative pose corrections directly from wrist-mounted RGB images. A coarse-to-fine two-stage control strategy is employed, and an image-pose paired dataset is constructed using ROS 2 and Isaac Sim. A convolutional neural network regresses relative translation and rotation from single-frame RGB inputs. In simulation, the planar positioning error decreases from 9.69 mm to 5.38 mm. On a real UR5e robot, the method achieves grasping success rates of 66.6% and 63.6% on two object categories and demonstrates robustness under rotational perturbations.

0 citationsRead paper

Decoding Market Emotion from Blockchain Activity: A Data-Driven Sentiment Classifier

Jul 16, 2026

This study addresses the challenge of deciphering Bitcoin market sentiment from multi-source heterogeneous data rather than directly forecasting prices. By integrating on-chain metrics, historical price data, and daily Twitter sentiment classifications, the authors construct a normalized dataset and employ an XGBoost model for sentiment classification. The work presents the first systematic fusion of on-chain, financial, and social media data specifically for sentiment interpretation in cryptocurrency markets. To enhance model interpretability, SHAP (SHapley Additive exPlanations) analysis is incorporated, revealing the pivotal role of on-chain features in sentiment determination. Experimental results demonstrate that the proposed approach achieves an average F1 score of approximately 0.84 in sentiment classification, thereby validating the efficacy and analytical insight gained through multi-source data integration in cryptocurrency sentiment analysis.

0 citationsRead paper

Improving Network Anomaly Detection via Choquet-Integral-Based Feature Aggregation

Jul 16, 2026

This study addresses the challenge of feature redundancy in high-dimensional network traffic data, which hinders the efficiency and real-time performance of anomaly detection. To overcome this limitation, the work proposes a novel feature aggregation framework that, for the first time, integrates the Choquet integral into network anomaly detection by combining adaptive weighting with incremental feature selection. This approach substantially reduces data dimensionality while simultaneously enhancing detection accuracy. Experimental results demonstrate that, using only the most critical features, the proposed method achieves up to a 7% improvement in accuracy over baseline models, compresses data volume by 77.5% (from 214 MB to 48 MB), and exhibits statistically significant superiority (p < 0.05).

0 citationsRead paper

AquaFeat+: an Underwater Vision Learning-based Enhancement Method for Object Detection, Classification, and Tracking

Dec 02, 2025International Conference on Advanced Robotics

Underwater videos are often degraded by insufficient lighting, color distortion, and turbidity, which significantly impair robotic perception performance. To address this challenge, this work proposes AquaFeat+, a task-oriented, plug-and-play, end-to-end enhancement framework that, for the first time, jointly optimizes underwater visual enhancement and downstream perception tasks—specifically designed for machine perception rather than human vision. AquaFeat+ integrates color correction, hierarchical feature enhancement, and an adaptive residual output module, with training directly guided by task-specific loss functions. Evaluated on the FishTrack23 dataset, AquaFeat+ substantially improves performance in object detection, classification, and tracking, demonstrating its effectiveness in enhancing underwater robotic perception.

0 citationsRead paper

SpellForger: Prompting Custom Spell Properties In-Game using BERT supervised-trained model

Nov 19, 2025

This work addresses the challenge of deeply integrating AI into core gameplay mechanics to enable real-time, natural-language-driven co-creation of personalized spells by players. We propose and implement SpellForger, a prototype game in which players describe desired spell effects using natural language; a fine-tuned BERT model parses these inputs in real time, followed by semantic mapping and parameter-balancing algorithms that generate numerically sound, mechanic-compatible abilities, executed instantaneously within Unity. To our knowledge, this is the first system to embed a pre-trained language model (BERT) directly into the game loop as a real-time gameplay generator, establishing natural language as a direct, interactive control modality. Experimental evaluation demonstrates effective synergy among semantic fidelity, numerical balance, and expressive player creativity—validating the feasibility and engagement of AI-augmented player–AI co-creation at the mechanistic level.

0 citationsRead paper

Recent publications

Latest Papers

Self Supervised Learning from Automatically Generated Demonstrations for Visual Robotic Manipulation

Aug 01, 2026

This work proposes a self-supervised visual manipulation method that eliminates the need for manual programming, human demonstrations, or extrinsic camera-robot calibration, which traditionally incur high deployment costs. The robot autonomously generates visual demonstrations near the target pose and learns relative pose corrections directly from wrist-mounted RGB images. A coarse-to-fine two-stage control strategy is employed, and an image-pose paired dataset is constructed using ROS 2 and Isaac Sim. A convolutional neural network regresses relative translation and rotation from single-frame RGB inputs. In simulation, the planar positioning error decreases from 9.69 mm to 5.38 mm. On a real UR5e robot, the method achieves grasping success rates of 66.6% and 63.6% on two object categories and demonstrates robustness under rotational perturbations.

0 citationsRead paper

Decoding Market Emotion from Blockchain Activity: A Data-Driven Sentiment Classifier

Jul 16, 2026

This study addresses the challenge of deciphering Bitcoin market sentiment from multi-source heterogeneous data rather than directly forecasting prices. By integrating on-chain metrics, historical price data, and daily Twitter sentiment classifications, the authors construct a normalized dataset and employ an XGBoost model for sentiment classification. The work presents the first systematic fusion of on-chain, financial, and social media data specifically for sentiment interpretation in cryptocurrency markets. To enhance model interpretability, SHAP (SHapley Additive exPlanations) analysis is incorporated, revealing the pivotal role of on-chain features in sentiment determination. Experimental results demonstrate that the proposed approach achieves an average F1 score of approximately 0.84 in sentiment classification, thereby validating the efficacy and analytical insight gained through multi-source data integration in cryptocurrency sentiment analysis.

0 citationsRead paper

Improving Network Anomaly Detection via Choquet-Integral-Based Feature Aggregation

Jul 16, 2026

This study addresses the challenge of feature redundancy in high-dimensional network traffic data, which hinders the efficiency and real-time performance of anomaly detection. To overcome this limitation, the work proposes a novel feature aggregation framework that, for the first time, integrates the Choquet integral into network anomaly detection by combining adaptive weighting with incremental feature selection. This approach substantially reduces data dimensionality while simultaneously enhancing detection accuracy. Experimental results demonstrate that, using only the most critical features, the proposed method achieves up to a 7% improvement in accuracy over baseline models, compresses data volume by 77.5% (from 214 MB to 48 MB), and exhibits statistically significant superiority (p < 0.05).

0 citationsRead paper

AquaFeat+: an Underwater Vision Learning-based Enhancement Method for Object Detection, Classification, and Tracking

Dec 02, 2025International Conference on Advanced Robotics

Underwater videos are often degraded by insufficient lighting, color distortion, and turbidity, which significantly impair robotic perception performance. To address this challenge, this work proposes AquaFeat+, a task-oriented, plug-and-play, end-to-end enhancement framework that, for the first time, jointly optimizes underwater visual enhancement and downstream perception tasks—specifically designed for machine perception rather than human vision. AquaFeat+ integrates color correction, hierarchical feature enhancement, and an adaptive residual output module, with training directly guided by task-specific loss functions. Evaluated on the FishTrack23 dataset, AquaFeat+ substantially improves performance in object detection, classification, and tracking, demonstrating its effectiveness in enhancing underwater robotic perception.

0 citationsRead paper

SpellForger: Prompting Custom Spell Properties In-Game using BERT supervised-trained model

Nov 19, 2025

This work addresses the challenge of deeply integrating AI into core gameplay mechanics to enable real-time, natural-language-driven co-creation of personalized spells by players. We propose and implement SpellForger, a prototype game in which players describe desired spell effects using natural language; a fine-tuned BERT model parses these inputs in real time, followed by semantic mapping and parameter-balancing algorithms that generate numerically sound, mechanic-compatible abilities, executed instantaneously within Unity. To our knowledge, this is the first system to embed a pre-trained language model (BERT) directly into the game loop as a real-time gameplay generator, establishing natural language as a direct, interactive control modality. Experimental evaluation demonstrates effective synergy among semantic fidelity, numerical balance, and expressive player creativity—validating the feasibility and engagement of AI-augmented player–AI co-creation at the mechanistic level.

0 citationsRead paper