TaijiFlow AI Research

Pose-Based AI for Continuous Human Movement Analysis
Ph.D. Doctoral Research in Progress (2026–2029)
Researcher: Weerayuth Uarjaipra
Affiliation: Multidisciplinary and Interdisciplinary School, Chiang Mai University, Thailand
Institutional Contact: weerayuth_u@cmu.ac.th
System Prototype: ai.taijiflow.org
Main Thai Research Portal: research.taijiflow.org

---

1. Executive Summary

TaijiFlow AI is an ongoing Ph.D. research effort by Weerayuth Uarjaipra, a doctoral student at the Multidisciplinary and Interdisciplinary School, Chiang Mai University, Thailand.

The project investigates pose-based computational movement intelligence, utilizing Chen-style Taijiquan (Silk Reeling Exercise / 缠丝劲) as an exemplary, highly demanding somatic domain. The core scientific question focuses on transforming tacit, continuous spiral human movement kinematics into computable, interpretable, and objective quality assessments. The project bridges biomechanics, computer vision, and machine learning, alongside developing an accessible browser-based system for real-time motor learning feedback.

---

2. Research Objectives

The research is organized around four primary scientific and engineering goals:

1. Continuous Spatiotemporal Representation: Modeling the interdependent kinetic chains between weight transference, pelvic rotation, spinal upright alignment, and bilateral arm coordination from temporal landmark sequences.
2. Graph and Attention-based Movement Quality Assessment: Benchmarking Spatio-Temporal Graph Convolutional Networks (ST-GCN) and Self-Attention Transformers against expert somatic ratings to evaluate movement fidelity and consistency.
3. Neuro-Symbolic Explainable AI (XAI): Combining data-driven model attention weights with verified biomechanical rule constraints (12 formalized Taiji principles) to generate pedagogical feedback that is physically meaningful, interpretable, and safe.
4. Accessible Client-Side Deployment: Evaluating runtime efficiency, numerical parity, and latencies when exporting models to ONNX Runtime Web (WebGPU / WebAssembly), minimizing cloud inference dependency while preserving user physiological privacy.

---

3. Current Academic Status & Phase 3 Roadmap

The research trajectory spans three progressive phases:

Phase	Timeframe	Research Scope & Focus	Milestone Status
Phase 1	2024–2025	Master's Independent Study (IS): Single-Hand Silk Reeling (4 forms), 9 biomechanical heuristics, software engineering standards (ISO/IEC 29110).	Completed
Phase 2	2025–2026	Funded research expanding to 12 standard forms (single & double hand), 12 universal rules, automated body calibration, and Supabase landmark data pipeline.	Current Baseline Release
Phase 3	2026–2029	Ph.D. Dissertation Research: ST-GCN / Transformer sequence modeling, Neuro-Symbolic XAI, ethical protocol design (IRB), and controlled efficacy study.	Current Work — In Progress

Academic Standing:

• Degree: Doctor of Philosophy (Ph.D.) in Multidisciplinary and Interdisciplinary Studies
• Institution: Chiang Mai University, Thailand
• Proposal Status: Proposal defense preparation in progress (Scope and experimental matrix subject to academic review committee).
• Ethics Review: Human-subject research protocol (IRB) is prepared separately prior to clinical cohort recruitment.

---

4. Research Infrastructure vs. Client Deployment

To avoid architectural ambiguity, TaijiFlow AI strictly delineates the Research Training Pipeline from the Client Inference Pipeline:

• Research Training: Executed in Python using PyTorch and NVIDIA CUDA on dedicated GPU infrastructure (high VRAM capacity required for multi-frame graph convolutions).
• Client Inference: Executed directly in the user's web browser via ONNX Runtime Web using WebGPU or WebAssembly fallback. No raw camera feed is transmitted to external servers.

---

5. Current Prototype & Limitations

An interactive prototype is accessible at ai.taijiflow.org.

NOTE

Prototype Scope Clarification:
The live web prototype demonstrates the Phase 1 & Phase 2 heuristic evaluation engine (MediaPipe Pose + rule-based geometric constraints). Deep learning models (ST-GCN and Transformer attention architectures) developed under Phase 3 are actively trained and evaluated within our laboratory environment and will be deployed upon experimental validation.

---

6. Research Metadata

Item	Details
Project	TaijiFlow AI (Phase 3)
Principal Researcher	Weerayuth Uarjaipra (Ph.D. Student)
Affiliation	Multidisciplinary and Interdisciplinary School, Chiang Mai University, Thailand
Current Phase	Phase 3 — Deep Learning & Continuous Movement Intelligence (In Progress)
Proposal Status	Defense preparation in progress
Target Period	2026–2029
Research Domains	Artificial Intelligence, Computer Vision, Spatio-Temporal Graph Neural Networks, Human Movement Analysis, Biomechanics
AI / Training Stack	Python, PyTorch, NVIDIA CUDA (ST-GCN, Spatio-Temporal Transformer)
Pose Pipeline	MediaPipe Pose (33 3D body landmarks)
Deployment Stack	ONNX Runtime Web, WebGPU / WebAssembly
Web Platforms	VitePress (Research Hub), Astro (Interactive Web System)
Institutional Contact	weerayuth_u@cmu.ac.th
Source Repository	GitHub: vyut/TaijiFlow
License	MIT License (Software) / CC BY-NC 4.0 (Research Artifacts)
Last Content Update	September 2026

---

7. Research Collaboration & Infrastructure Support

We actively welcome discussions with academic labs, technology partners, and computing programs in the following focused areas:

• Academic Hardware & Compute Access: Seeking evaluation access or academic pricing for CUDA-compatible computing nodes (e.g., NVIDIA workstations, HPC cluster allocation, or cloud GPU credits) for ST-GCN and Transformer ablation experiments.
• Biomechanical Reference Validation: Cross-validation collaboration with sports science or biomechanical motion laboratories (MoCap, 3D kinematic benchmark subsets).
• Expert Somatic Calibration: Experienced Taijiquan masters and somatic instructors interested in reviewing standardized movement scoring rubrics.

Inquiries regarding academic collaboration should be directed to the researcher via institutional email: weerayuth_u@cmu.ac.th.

---

8. Citation

@phdthesis{uarjaipra2026taijiflow,
  author       = {Weerayuth Uarjaipra},
  title        = {TaijiFlow AI: Pose-Based Artificial Intelligence for Continuous Human Movement Analysis and Feedback},
  school       = {Multidisciplinary and Interdisciplinary School, Chiang Mai University},
  year         = {2026},
  address      = {Chiang Mai, Thailand},
  url          = {https://research.taijiflow.org/en/}
}