I am Ruqian Xie, an AI and robotics developer at Beijing Information Science and Technology University. I build verifiable intelligent systems at the intersection of multi-agent collaboration, generative AI, and humanoid motion learning.
My work follows one recurring question: how can generated content and human motion become systems that are observable, reviewable, and reliable enough to use?
PoetryEduAgent · PoetryVision · Humanoid Motion Intelligence
A multi-agent learning system that turns classical Chinese poetry into personalized teaching resources, generated imagery, and assessable student experiences.
PoetryEduAgent coordinates specialized agents for knowledge retrieval, poem analysis, learning-resource generation, text-to-image prompting, visual generation, and assessment. Text and imagery pass through independent reviewers before a final dual-review gate, while teachers can inspect the execution trace and request targeted revisions.
- Personalized learning: adapts explanations, imagery, activities, and assessments to grade level and learning goals.
- Independent review paths: evaluates educational content and generated images separately before release.
- Observable orchestration: records agent events, review decisions, correction attempts, and learning reports.
Built with: Qwen · Qwen-VL · DeepSeek · Kolors · RAG · FastAPI · SQLite · Server-Sent Events
Explore PoetryEduAgent → · View the companion PoetryVision project →
View the multi-agent execution workspace
An end-to-end research workflow for converting human motion into validated trajectories and motion-imitation policies for an Omni 29-DoF humanoid.
Rather than treating retargeting, simulation, and reinforcement learning as separate experiments, I connect them as one continuous engineering thread:
Human motion → GMR / inverse-kinematics retargeting → MuJoCo and Blender acceptance → Isaac Lab / UniLab motion-imitation training → checkpoint, ONNX, and Sim2Sim verification
- Motion representation: skeleton topology, frame rate, coordinates, root semantics, quaternion conventions, and 29-joint policy order.
- Physical acceptance: joint limits, continuity, contacts, collision behavior, root motion, and visual fidelity across long sequences.
- Learning and delivery: PPO-based imitation, observation-contract checks, checkpoint compatibility, policy export, and simulation-first deployment boundaries.
Research stack: Python · PyTorch · MuJoCo · Blender · Isaac Lab · UniLab · RSL-RL · PPO · ONNX
Capital College Students Entrepreneurship Plan Competition · Participant
Developed PoetryEduAgent as a competition-oriented AI education system combining multi-agent collaboration, classical-poetry understanding, text-to-image generation, independent text and visual review, and student assessment.
Official competition announcement →
Second Edition · Beijing · Participant
Working on humanoid motion intelligence for competition preparation, including full-body motion retargeting, long-sequence visual acceptance, trajectory preparation, and motion-imitation research.
AI systems — multi-agent orchestration, model routing, retrieval-augmented generation, multimodal review, bounded correction, and traceable execution.
Embodied intelligence — human-motion retargeting, whole-body control, simulation validation, reinforcement learning, and policy delivery.
Engineering — Python, FastAPI, PyTorch, SQLite, Linux, Docker, Git, FFmpeg, and reproducible evaluation workflows.