Jisheng Qin

Jisheng Qin (秦吉胜)

Ph.D

Graph Neural Network; Large Language Models; Computer Human Interaction

Email: jisheng.qin.vip@gmail.com

[Google Scholar]    [GitHub]

The goal is to create a universal behavior tree model that enhances the performance of various downstream tasks in robotics. Starting with health monitoring tasks in real-world elderly care scenarios, we analyze the execution patterns of service robots' behavior trees. By modeling task dependencies and logical connections, we employ self-supervised learning methods to capture spatiotemporal correlations, facilitating task logic transferability across different care environments.

We pre-trained a foundational behavior tree model with 500 million parameters using large-scale elderly care task data. Incorporating more complex behavior data, such as multi-robot collaboration and emergency scheduling, we expanded the model to 1 billion parameters. The enhanced model supports dynamic task allocation, emotional companionship, and rapid adaptation to diverse scenarios through a unified framework called CareAlign. To address data scarcity, we developed BehaviorSynth, a diffusion-based model for generating synthetic behavior tree data, enabling robust generalization and scalability.

The goal is to establish a large-scale foundational model for dynamic skeletal action recognition aimed at advancing intelligent health monitoring and behavior analysis in imaginative elderly care scenarios. We aim to establish a foundational model for dynamic skeletal action recognition, enhancing intelligent health monitoring and behavior analysis in elderly care. This model enables tasks such as fall detection, daily activity monitoring, and motion rehabilitation, promoting personalized health management for older adults. Starting with real-world scenarios, we analyze skeletal motion patterns, identifying actions while monitoring health in real time.

Utilizing biomechanics and graph neural networks, we built a skeletal keypoint data model and applied self-supervised learning to capture spatiotemporal dependencies. Pretrained on extensive skeletal datasets, our model, with up to 1 billion parameters, adapts to various downstream tasks and environments. The BoneAlign framework enhances cross-scenario adaptability with minimal labeled data, supporting anomaly detection and health recommendations. To mitigate data scarcity, we developed ActionSynth to generate diverse dynamic skeletal datasets, ensuring robustness and scalability in real-world applications.

The goal is to pre-train a general-purpose graph foundation model using extensive graph data. After appropriate fine-tuning, this model aims to achieve satisfactory performance across a variety of downstream tasks, showcasing its broad application potential while addressing numerous challenges.

We aim to pretrain a general-purpose graph foundation model using extensive graph data. This model delivers robust performance across a variety of downstream tasks after fine-tuning, addressing challenges and demonstrating broad application potential. Our graph foundation model features enhanced representational capabilities, leveraging advanced pretraining strategies and self-supervised learning methods to capture graph structures effectively.

During adaptation to downstream tasks, we explored intrinsic factors affecting model performance and designed parameter-efficient techniques. To address dataset scarcity, we introduced DGraph, a large-scale dynamic graph action recognition dataset. This resource supports the development of robust graph-based models, advancing applications in areas like link prediction, recommendation systems, and dynamic network analysis.