Building Human-aware Embodied Agent Systems

EEG & Multimodal Sensing → State Modeling → Decision → Robot Action

English · 简体中文

🌐 Personal Website

I am currently building human-aware embodied agent systems that integrate multimodal sensing, state modeling, decision-making, and action. My work focuses on AI systems combining EEG, vision, and physiological or behavioral signals, with robotics integration on embodied platforms such as ALOHA, Astribot, and ZsiBot.

These systems follow closed-loop architectures: perception → state estimation → decision → action execution → feedback.

• EEG → Intention Decoding → Decision → STM32 Vehicle Control
• Facial Emotion → Decision → Safety-Gated Astribot Action
• EMG Gesture → Decision → ZsiBot Robot Control
• Vision + Robot State + EEG Signals → OpenPI Policy → ALOHA Dual-Arm Action

My research interests include cross-subject EEG generalization, multimodal fusion across EEG, vision, language, and audio, as well as behavior-driven human state modeling beyond traditional BCI settings. I am also interested in deploying real-time AI for human-aware embodied agent systems.

• Multifractal + Graph-based + Transformer-based model
• Cross-subject generalization on SEED-VII dataset
• Focus on robustness and generalization
• Paper: Local-Global Feature Fusion for Subject-Independent EEG Emotion Recognition
• Accepted for Oral Presentation at IEEE EMBC 2026

• End-to-end EEG intention recognition system
• Model training (LSTM / SVM / etc.) + real-time control
• Integrated with embedded system (STM32 + Bluetooth)
• Honored as The 13th Cloud Programming World Cup - first prize

Demo video: coming soon

• Multimodal human-state sensing with facial emotion recognition, EEG signals, and BCI paradigms
• Built a DeepFace-based emotion recognition prototype for real-time robot interaction — notes
• Explored SSVEP-based robot control and motor-imagery classification with EEGNet
• Closed-loop system (perception → state estimation → decision → action → feedback) with stable event-triggering logic to reduce noisy or unstable robot actions
• π0.5 VLA model fine-tuning on AgileX Aloha for bimanual manipulation — notes
• OpenPI policy serving and dual-arm Piper inference with failure recovery — notes
• EMG gesture recognition controlling ZsiBot ZSL-1W wheeled-legged robot (~40ms latency) — notes

• 42-subject EEG-Audio-Video dataset with leakage-free splits; built complete unimodal baselines and late fusion achieving 0.5729 accuracy

• Docker-first personal website with Caddy reverse proxy and WordPress
• Migrated to AI-ready infrastructure with FastAPI and future agent services
• Notes: Amor Fati AI Infrastructure

• Programming: Python, C/C++, Java, SQL, JavaScript, Shell

• AI / Machine Learning: PyTorch, TensorFlow, scikit-learn, OpenCV, MNE, Braindecode, EEGNet, Transformer Models, NumPy, Pandas, Matplotlib

• AI Engineering Workflow: Claude Code, OpenAI Codex

• Systems & Infrastructure: Linux, Docker, Kubernetes, Git, SSH, VSCode Remote, MySQL, Jupyter, LaTeX

• Robotics & Embedded: Astribot SDK, ALOHA, OpenPI, STM32 Embedded Control

I am open to collaboration and discussion on multimodal agent systems, embodied AI, EEG and vision fusion, human state modeling, scalable agent architectures, and real-world deployment of intelligent systems.

• Email: daniel.zhengzhou@gmail.com
• LinkedIn: www.linkedin.com/in/zheng-zhou-cs

Always learning to balance.