whoami

你好，我是 Ningwei Bai。

中文简介

我是来自 University College London 的 Robotics and AI (MEng) 四年级学生，主要关注机器人、强化学习、视觉语言动作模型和大模型后训练。这个博客像是一个个人工作台：我会在这里记录研究、学习笔记、工具折腾、项目想法，以及那些还在路上的思考。

研究之外，我也很喜欢咖啡，持有 SCA Intermediate Barista 证书。对我来说，写代码、做实验和冲咖啡有点像：都需要耐心、手感，以及对细节保持一点偏执。

English Bio

I am Ningwei Bai, a fourth-year Robotics and AI (MEng) student at University College London. My interests sit around robotics, reinforcement learning, vision-language-action models, and post-training for reasoning large language models.

Outside research, I am also a coffee person with an SCA Intermediate Barista certificate. I like building things carefully, whether that means a robot policy, a small web tool, or a cup of coffee that actually tastes like the idea in my head.

CV Snapshot

Education

  • University College London, Robotics and AI (MEng), London, UK Sep 2023 - Jun 2027 expected; integrated undergraduate and master's programme; average score 79.5%.

Research Interests

  • Reinforcement learning for robotics and nonlinear control systems.
  • Vision-language-action models for robotic manipulation.
  • Adaptive dynamic programming, reasoning LLM post-training, and multimodal retrieval.

Selected Publications

  • SignVLA: real-time sign language-guided robotic manipulation via Attention LSTM and VLA models. ICAC 2026, accepted.
  • An Event-Triggered Robust Reinforcement Learning Scheme for Nonlinear Systems with Unknown Disturbances. CCC 2026, accepted.
  • Reinforcement learning based optimal control: a survey of ADP for manipulators and wheeled mobile robots. Advanced Mechatronics, accepted.
  • Zero-shot Decomposed Retrieval Enhancement for Visually Rich Document via Broad Search and Deep Reasoning.
  • Post-Training for Reasoning LLMs with Reinforcement Learning: A Stability-Efficiency Perspective. ICAC 2026, accepted.

Research Experience

  • Vision-Language-Action Model Control by Sign Language Deployed NVIDIA GR00T on a robotic manipulator, built a MediaPipe Hands gesture feature pipeline, and connected sign language understanding to VLA model commands.
  • Reinforcement Learning for Nonlinear Control Systems Combined event-triggered mechanisms, extended state observers, and adaptive dynamic programming to reduce redundant updates while preserving control performance.
  • Visual RAG Systems Designed a zero-shot multimodal query decomposition method for visually rich document retrieval, improving Recall@1 in experiments.

Projects

  • BabyBench Infant Self-Touch RL: built a Mujoco-based self-touch learning environment and trained early exploration behavior with PPO.
  • Emotion-Aware AI Intelligent Photo Album: fine-tuned dialogue and expression-recognition components with Qwen, YOLO, ResNet50, and Gradio.
  • 7-DOF Robotic Arm Simulation: a self-recorded course on MATLAB / Simulink-based 7-DOF robotic arm simulation, covering kinematics, dynamics, control-system design, and modelling practice.

Skills

Python, PyTorch, C, C++, Matlab, Simulink, ROS, ROS2, Linux, STM32 HAL, Solidworks, Fusion 360, Docker.

Download CV