Welcome, I am Wentao Guo. Building embodied intelligence through
data loops and robot hands.

Designed and standardized a task-specific action workflow, collected UMI data, supplemented targeted data with DAgger, and fine-tuned the model for robust deformable-object manipulation.

Achieved vertical adaptability via linear linkages, enabling pinch-scoop switching under environmental constraints. Fabricated six 3D-printed prototypes for systematic validation.

Modular design with navigation, grasping, perception, monitoring, cleaning, and dialogue. Supports modular camera setup with AI recognition. Implemented DQN habit-learning and semi-autonomous execution on Jetson Nano.

Meeting agent built on Jiuge on-device LLM: integrates local DB, custom summarization, and ASR, with three RAG optimizations (query planning, confidence gating, adversarial robustness) for better minutes and retrieval.

For distribution grids under extreme weather, we integrate strong AI with digital twins: time-series models and local knowledge base for early warnings, vulnerability assessment, and damage location prediction; an industry LLM connects multimodal monitoring with operations research for proactive monitor-alert-dispatch-recovery; a digital-twin platform provides panoramic visualization and strategy simulation.

Sleep-focused pipeline: from staging and scoring to LLM analysis and fine-tuning. Applied fine-tuning and constraints to reduce bias, with clustering-based updates planned for continuous improvement.