Enabling Direct Robot Learning from Human Demonstrations for Robotic Ultrasound Scanning: Design of an Instrumented Probe Attachment

Institute

Chair for Computer Aided Medical Procedures & Augmented Reality (TUM-CIT)

Type

Semester Thesis / Master's Thesis /

Content

experimental / theoretical / constructive /  

Description

Background and Motivations:

Ultrasound scanning is widely used but strongly depends on the operator’s skill to find correct views while maintaining stable contact and appropriate pressure. This variability limits repeatability and makes it difficult to scale ultrasound examinations across users and sites. Robot learning from human demonstrations is a promising way to teach robots ultrasound scanning behaviors directly from expert operation, but it requires high-quality demonstration data. In particular, learning robust policies needs synchronized measurements of ultrasound images, accurate 6-DoF probe pose, and contact force. Such multimodal recordings are often not readily available with standard clinical probes. This thesis addresses this gap by developing an instrumented probe attachment that enables external optical tracking and force measurement during manual scanning, and by using the recorded data to train a model for automatic ultrasound scanning tasks.

 

Research Objectives:

Hardware design: Design and prototype an instrumented probe attachment that enables external optical tracking and force measurement during ultrasound scanning.
Data collection: Build a synchronized recording pipeline and collect human demonstrations (ultrasound images + probe pose + force) for one or more scanning tasks.
Robot learning: Train and evaluate a learning-based model that learns the scanning behavior from demonstrations and performs the task automatically.

Requirements

Prerequisites:

Interests in Robotics, Computer Vision, and Deep Learning.
Experience in Mechanical Design and Building Prototypes.
Good experience in Pytorch and Robotic Control (ROS and Ubuntu experience will be a plus)
Prior Knowledge about medical imaging (not necessary).

 

What We Offer:

You will gain hands-on experience with real Franka or KUKA robotic arm, advanced medical imaging systems and computer vision sensors. The project is expected to start as soon as possible, and depending on the progress, there will be a strong opportunity to submit the work to the upcoming robotic conferences or journals.

 

Application:

Interested applicants please send the brief CV, academic transcripts via email M.Sc. Yuan Bi  (yuan.bi@tum.de) or M.Sc. Xuesong Li (xuesong.li@tum.de)

Possible start

sofort

Contact

Xuesong Li
Xuesong.Litum.de