Exploring the Cognitive Architecture
Of Embodied Agents.
Undergraduate Student @ Xi’an University of Architecture and Technology (XAUAT).
Working on Whole-body Control and World Models for Humanoid Robots.
My research seeks to bridge the gap between high-level reasoning and low-level torque control. I am less interested in the hardware Itself, but rather in how we can instill biological plausibility and adaptive intelligence into these iron vessels.
探索具身智能体的
认知架构
本科生 @ 西安建筑科技大学。
致力于人形机器人的 全身控制 (Whole-body Control) 与 世界模型 (World Models) 研究。
我的研究旨在弥合“高层认知推理”与“底层力矩控制”之间的鸿沟。相比于单纯的硬件本体，我更着迷于如何赋予这些钢铁躯壳以生物合理性和自适应智能。
Selected Research & Manuscripts 精选研究与手稿
Proposed a hierarchical framework where a Transformer-based world model generates latent trajectories, which are then tracked by a convex MPC controller. This approach allows the humanoid to "Imagine" physics before stepping.
提出了一种分层架构：利用基于 Transformer 的世界模型生成潜在轨迹，并通过凸优化 MPC 控制器进行跟踪。该方法使人形机器人能够在迈步前先“想象”物理后果，从而实现更加稳健的动态行走。
Implemented PPO from scratch to train a humanoid agent in MuJoCo. Analyzed the impact of reward shaping on gait symmetry and energy efficiency.
从零实现了 PPO 算法并在 MuJoCo 环境中训练人形智能体。重点分析了不同奖励函数设计（Reward Shaping）对步态对称性及能量效率的具体影响。
Talks & Notes 演讲与笔记
I document my learning path and theoretical derivations. 记录我的学习路径与理论推导。
-
Oct 2024The Geometry of Control: Lie Algebra in Robotics 控制的几何学：机器人学中的李代数
Seminar talk at Robotics Group. Explaining SO(3) and SE(3) manifolds. 机器人组内研讨会分享。解析 SO(3) 与 SE(3) 流形基础。
[Slides]
Background 背景与技能
Advanced Coursework 核心课程
- Convex Optimization
- MIT 6.8210：Underactuated Robotics
- Deep Learning & Generative Models
- CS285:Deep Reinforcement Learning
- CMU 16-831:Introduction to Robot Learning
- 凸优化 (Convex Optimization)
- 最优控制理论 (Optimal Control)
- 深度学习与生成式模型
- CS285: Deep Reinforcement Learning (UC Berkeley)
Technical Stack 技术栈
Frameworks: PyTorch, JAX
Simulation: MuJoCo, Isaac Gym