Developed remote teleoperation system for bimanual dexterous manipulation and data collection using PyMyCobot. Implemented real-time control protocols and collected demonstration data for imitation learning tasks like pick and place, enabling precise manipulation through intuitive human-robot interfaces.
Python
ROS
MyCobot
Teleoperation
Engineered comprehensive data extraction pipeline for UMI, implementing SLAM-ORB3 for trajectory extraction from GoPro footage and gripper state estimation. Developed custom Rerun visualizations for real-time 7-DOF trajectory analysis (3D position + 4D quaternion), synchronizing visual SLAM outputs with gripper sensor data. Pipeline processes raw demonstrations into training-ready datasets with precise temporal alignment between camera poses and manipulation actions.
SLAM-ORB3
Rerun
Python
OpenCV
7-DOF Tracking
Developed annotation tools for robotic manipulation datasets using embodied Chain-of-Thought methods. Created interfaces for labeling gripper poses, contact points, and task-relevant keyframes in demonstration data.
Python
Computer Vision
Annotation Tools
PyTorch
Architected scalable data storage and versioning system for robotics datasets. Deployed MinIO for S3-compatible object storage with LakeFS for Git-like version control of large sensor data, enabling reproducible ML experiments and efficient data management across distributed teams.
MinIO
LakeFS
S3 API
Docker
Data Versioning
Built end-to-end robotic learning platform using LeRobot framework. Implemented data collection pipelines, teleoperation interfaces, and policy training infrastructure for the ISO-101 robot, enabling rapid prototyping of manipulation behaviors from human demonstrations.
LeRobot
PyTorch
ROS2
Imitation Learning