Projects &
Capabilities项目与能力

For my undergraduate thesis, I independently designed an integrated humanoid robot joint combining a permanent-magnet synchronous motor, a single-stage planetary gearbox, and drive and sensing components. I compared winding configurations and materials through finite-element analysis, then optimized motor parameters using sensitivity analysis, response-surface models, and a genetic algorithm. Simulation showed about 15% lower peak-to-peak cogging torque and 5% higher average output torque. I completed electromagnetic–thermal analysis, efficiency and loss maps, SolidWorks assembly design, and structural strength checks for a 6:1 quasi-direct-drive joint.

本科毕业设计中，独立完成一体化人形机器人关节方案，将永磁同步电机、一级行星减速器及驱控传感部件集成设计。通过有限元分析比较电机槽极配合与材料，结合灵敏度分析、响应面模型和遗传算法优化电机参数；仿真中齿槽转矩峰峰值降低约 15%，平均输出转矩提高约 5%。完成电磁热耦合分析、效率与损耗图谱、SolidWorks 总装设计及传动结构强度校核，形成减速比 6:1 的准直驱关节方案。

I designed and developed a haptic gripper trigger for Franka Panda and integrated it into the teleoperation system. I implemented encoder readout, motor FOC control, and the force-feedback interface so the operator can control the gripper and feel grasp feedback through the trigger. The ESP32 firmware uses SimpleFOC, supports position and torque modes, and communicates with the robot through UART and ROS.

设计并开发了用于控制 Franka Panda 夹爪的力反馈夹爪 Trigger，并接入遥操作系统。独立完成编码器读取、电机 FOC 控制与力反馈接口，使操作者能够通过 Trigger 控制夹爪，并感受抓取反馈。固件基于 ESP32 和 SimpleFOC，支持位置与力矩模式切换，通过 UART 和 ROS 与机器人通信。

I study how fixed-impedance teleoperation demonstrations can become continuous variable-impedance policies for learning. CMDIR introduces a continuous task-manifold impedance representation and Quality-to-Fast optimization to adjust contact behavior while preserving task response, then uses the optimized controllers to supervise imitation learning. I implemented demonstration retargeting, policy training, and robot execution, with experiments on planar wiping, curved wiping, and pushing.

研究如何把固定阻抗遥操作示教转化为可用于学习的连续变阻抗控制策略。提出连续任务流形阻抗表示和 Quality-to-Fast 优化方法，在保持任务响应的同时调整接触行为，并用优化后的控制器监督模仿学习。完成从示教重定向、策略训练到机器人执行的实现，在平面擦拭、曲面擦拭和推动任务中开展实验。

I developed task-manifold impedance retargeting to turn fixed-impedance demonstrations into variable-impedance controllers for contact tasks. The method separates work, exertion, and support requirements to moderate contact while preserving task response. I implemented the method and evaluated it on Franka Panda in planar wiping, pick-and-place, and pushing tasks.

提出任务流形阻抗重定向方法，将固定阻抗示教转化为适合接触任务的变阻抗控制器。方法按工作、施力和支撑的不同需求组织控制，在保留任务响应的同时减缓接触冲击。完成方法实现及 Panda 实机实验，验证于平面擦拭、抓放和推动任务。

I developed a Cartesian variable-impedance controller for Franka Panda, used for force-feedback teleoperation, demonstration replay, and learned-policy execution. It supports task-space stiffness and damping adjustment, null-space posture control, torque-rate limiting, and synchronized pose and impedance commands across contact-manipulation experiments.

为 Franka Panda 开发笛卡尔变阻抗控制器，用于力反馈遥操作、示教回放和学习策略执行。实现任务空间刚度与阻尼调节、零空间姿态控制和力矩变化率限制，支持位姿与阻抗同步下发，让同一套控制器服务于不同接触操作实验。

I study phase transitions and stalled motion in action-chunk prediction for multi-stage contact-rich manipulation. TRACT routes predicted actions by task phase and uses a response-deficit integrator to compensate for motion suppressed by contact. I implemented training, inference, and hardware deployment and evaluated the method in multi-stage wiping.

面向多阶段富接触操作，研究动作块预测中的阶段切换与接触受阻问题。提出按任务阶段路由动作的模仿学习策略，并结合响应不足积分器补偿接触中的运动受阻。完成模型训练、推理与实机部署，在多阶段擦拭任务中验证。

During my exchange at the Smart Robot Design Lab, Tohoku University, I used YOLOv8, a purpose-built dataset, and a coupled two-model approach to extract garment hem points for robotic visual grasping.

在东北大学 Smart Robot Design Lab 交换期间，基于 YOLOv8、自建数据集与双模型结合方法提取衣物下摆特征点，用于机器人衣物视觉抓取。

I studied how load fluctuations in suspended backpacks affect human walking stability and energy expenditure, contributing experimental data analysis and quantitative dynamics modeling. The work was published in Mechanical Systems and Signal Processing and informs backpack performance evaluation and design optimization.

研究悬浮背包的负载波动如何影响人体步行稳定性与能耗，承担实验数据分析与动力学定量建模。成果发表于 Mechanical Systems and Signal Processing，为悬浮背包的性能评价与优化设计提供依据。

I built and continue to maintain a robot platform that reliably collects force-feedback teleoperation demonstrations for contact-rich tasks. It combines Franka Panda, Geomagic Touch, and a custom haptic gripper trigger for coordinated arm and gripper operation. I implemented hand–eye and force-sensor payload/bias calibration, with synchronized multi-view images, robot states, and contact forces for control and imitation-learning research.

自主搭建并长期维护力反馈遥操作机器人平台，稳定采集富接触任务的示教数据。平台以 Franka Panda 和 Geomagic Touch 为主体，接入自研力反馈夹爪 Trigger，实现机械臂与夹爪的协同遥操作。完成手眼标定、力传感器负载与零偏标定，并同步记录多视角图像、机器人状态和接触力，用于控制与模仿学习研究。

I developed a common deployment framework for running different robot learning models on the same Panda hardware and simulation platforms. It integrates ACT, Flow Matching, PI0.5, and my TRACT and CMDIR policies, with pose and variable-impedance control. The framework handles observation synchronization, inference, and action execution; demonstration conversion and run logging support the move from trained models to experiments.

自主开发通用策略部署框架，让不同的机器人学习模型能够在同一套 Panda 实机和仿真平台上运行。已接入 ACT、Flow Matching、PI0.5，以及自研的 TRACT 和 CMDIR 策略，支持位姿与变阻抗控制。框架负责观测同步、模型推理和动作执行，并配套示教数据转换与运行记录，便于训练后的模型直接进入实验。

I built an Isaac Lab platform for developing and evaluating Panda controllers and learning policies in contact-rich tasks. It includes pick-and-place, planar and curved wiping, and pushing, with impedance control and policy deployment corresponding to the hardware setup. Demonstration recording, physical replay, and perturbation experiments support repeated comparisons under different contact conditions.

基于 Isaac Lab 搭建 Panda 富接触操作仿真平台，用于控制器和学习策略的开发与评估。实现抓放、平面与曲面擦拭、推动等任务，接入与实机对应的阻抗控制和策略部署流程。支持示教录制、物理回放及扰动实验，便于反复比较机器人在不同接触条件下的表现。