About Me

Hello👋, I’m Tengfei Li ( AKA. 李腾飞). I am a three-year Master’s candidate in Artificial Intelligence at Sichuan University, building upon my undergraduate foundation in IoT Engineering at Zhengzhou University (a 211 Project institution). My research focuses on Human-Centric AI, specifically advancing Motion Understanding and Generation for digital human and game applications.

My research intersects Machine Learning, Computer Vision, Multimodal, Meta Human and AIGC, with active exploration of:

Text-driven real-time digital human interaction frameworks
Cross-modal alignment with uncertainty estimation techniques for pose reconstruction
High accuracy video motion capture for biomechanics analysis

I’m passionate about developing AI systems that bridge virtual and physical movement, with publications and projects addressing occlusion-robust pose analysis and synthesis.

For detailed credentials including publications, projects and demos, please visit my portfolio: CV Page. For convenience, I copy it below.

🎯 Research Focus

My core research agenda centers on human-centric artificial intelligence, with specialized expertise in:

🕴️ Human Pose Estimation: 3D Reconstruction Occlusion Handling MultiModel Methods
💃 Motion Synthesis & Generation: Interactive Text-Driven Control Long-Sequence Generation
👟 Video Motion Capture System: Multi-View High Accuracy Biomechanical Analysis

⚙️ Technical Arsenal

Meta Human: Stable Diffusion DiT VAE PPO Controllable
3D Vision: Mesh Motion Synthesis 3D Pose Reconstruction
Motion Capture: Multi-View Marker-less Biomechanics Analysis
Multimodal: Image Text Motion
Animation: Blender IK/FK Rotation Ragging
Optimization: KV Cache LoRA Noise Strategy CFG

🚀 Key Projects

💃 Self-Forcing Autoregressive Diffusion for Real-Time Text-Driven Motion Generation

AutoRegress Diffusion DiT Reinforcement Learning (PPO) Real-time Control

✅ SOTA Performance: Achieved 300+ FPS with KV Cache acceleration
✅ Enabled continuous long-sequence generation via novel self-forcing paradigm
✅ Balanced quality/speed via noise scheduling & annealing strategies

Real-Time Video Demo

Technical Note: This implementation is based on DART . Compared to DART, our method achieves more stable results through wider receptive fields while maintaining comparable frame rates.

🌐 MetaPose: Multimodal Enhancement and Transformation Alignment 3D Pose Restruction

Vision-Text-Spatial Fusion Cross-Modal Alignment Occlusion Robustness

✅ IEEE TMM (Q1 Under Review)
✅ SOTA Accuracy: Outperformed video-based methods using single-frame input
✅ Novel Framework: A unified distribution space for anatomical/textual/visual features

Video Demo in the Wild

👟 Multi-View Video Capture for Biomechanics Analysis

ViT Temporal Processing Triangulation

✅ EAAI (Q1, Major Revision)
✅Journal of Biomechanics (Top Journal of Orthopedics, Major Revision)
✅ High Accuracy: achieve the requirements of clinical biomechanical analysis.It is more accurate than the online markerless motion capture systems available on the market.

🤖 Provably Safe Vision-Based Teleoperation for Dual-Arm Surgical Robots

Surgical Robot Gesture Control Teleoperation

✅ IEEE ROBIO 2025

Demonstration

⚡ Connection-Aware Graph Convolution Networks

GCN Anatomical/Kinematic Modeling Multi-Level Aggregation

✅ HCIS Journal (JCR Q1, Under Review)
✅ 40% Faster than video-based approaches while maintaining SOTA accuracy

🩺 Reinforce Learning Enhanced CTA for Noninvasive Prediction

PPO CTA Medical Image Analysis

✅ Liver International (JCR Q1)
✅ High accuracy rate of identifying the ROI area of extremely severe lesions

🏆 Recognitions

National Encouragement Scholarship 2020-2022 (Top 10%)
4 SCI Q1 Papers (First-Author, 3 Under Review Now)
GPA 3.5+ @ Project 985 universitiy
GPA Top 10% @ Project 211 university

📜 Publications

Accept: Liver International (Q1) ROBIO 2025🥰
Major Revision: EAAI (Q1), Journal of Biomechanics (Q2)，
Under Review: HCIS (Q1), TMM (Q1), AAAI 2026 (CCFA)
Prepare to Submit: CVPR 2026 (CCFA)