Detailed project portfolio and technical highlights.
Deployment of a whole-body Control Algorithm for Humanoid Robots
Apr 10, 2026 - May 07, 2026
- Motion retargeting: Performed secondary development based on the open-source GMR toolkit, completed kinematic adaptation for Walker Tienkung (1.73m, 65kg, 30-DoF), and successfully converted human motion capture data in npz, bvh, and fbx formats into executable pkl joint trajectories for the robot.
- Reinforcement learning training: Built a distributed parallel simulation environment with Isaac Lab/Gym and trained policies with PPO and AMP. Iteratively optimized reward design based on reward curve analysis, and validated Sim2Sim policy transfer in a high-fidelity MuJoCo environment.
- Real-robot deployment (Sim2Real): To address dynamics gaps on hardware, applied domain randomization and delayed-signal enhancement for improved robustness, and ultimately achieved zero-shot trajectory tracking and stable walking on the Walker Tienkung robot.
- Paper writing: Based on this project, summarized the research outcomes and prepared the paper ATOM: A Training Framework for Humanoid Behavior Foundation Models with Imperfect Data, currently submitted to NeurIPS 2026 (under review).
Humanoid Robot Music Conductor
May 10, 2026 - May 21, 2026
- Motion retargeting and denoising: Used GVHMR to extract human poses from symphony conducting videos. For temporal noise in dual-arm trajectories, applied filtering for smoothing, then remapped motion to humanoid joint space via inverse kinematics.
- Upper-body control model training: Built a parallel simulation environment in Isaac Lab and trained a stable upper-body control policy with reinforcement learning, enabling stable walking together with simple and reliable upper-limb motions.
- Control system deployment: Designed a fused architecture for dual-arm control and lower-body balance, and fed smoothed target joint angles into the low-level PD controller at 50Hz, achieving high-fidelity, jitter-free conducting motions on real hardware.
YOLOv8 and MoveIt 2 based Object Pick-and-Place
Oct 16, 2025 - Dec 29, 2025
- Project overview: Built a full object pick-and-place pipeline for a humanoid robot in Isaac Sim, using YOLOv8 for object detection and MoveIt 2 for dual-arm motion planning. The system completed grasping through inverse kinematics in simulation and was finally deployed on real hardware to grasp specified objects.
- High-fidelity simulation platform: Created a realistic Isaac Sim environment, configured USD assets, and provided the base simulation platform for algorithm validation. Implemented ROS2 SDK interfaces consistent with the real robot and used ZMQ to bridge ROS and Isaac Sim communication.
- VLM deployment and object recognition: Deployed a VLM on the robot body so the robot could identify target objects, and used camera calibration to perform accurate TF transformations from perception results to the target grasp pose.
- MoveIt 2 configuration and IK solving: Configured MoveIt 2 for the humanoid robot dual arms, enabling the robot to receive Cartesian-space grasp targets, convert them into joint-space commands through inverse kinematics, and execute correct grasping motions through ROS interfaces.
Simulation and Control of the TIAGo Mobile Manipulator in Gazebo
May 16, 2025 - May 29, 2025
- Mobile manipulation planning: Implemented a full pick-and-place pipeline for the TIAGo mobile manipulator in Gazebo using ROS and MoveIt!.
- Motion planning and arm control: Configured the OMPL planner library (RRT and RRT*) for joint-space path planning and designed a custom time-parameterization method to smooth trajectory velocity.
- Autonomous navigation and obstacle avoidance: Built a navigation stack with Gmapping and Move Base, and combined it with ArUco pose feedback to achieve precise end-effector alignment and placement.