Master's thesis - Reinforcement‑learning‑based teleoperation for humanoid control using a motion capture suit
30.07.2025, Abschlussarbeiten, Bachelor- und Masterarbeiten
Join our cutting-edge research on humanoid teleoperation and develop a unified RL+BC controller to bring real human motion to life on our Unitree G1 robot.
Background & Motivation
Within our project team at the Chair of Microtechnology and Medical Device Technology (MiMed) we are researching the teleoperation of humanoid robots. As part of this research, a sensor suit has been developed that continuously records and provides kinematic data for the entire human body. Teleoperation allows humans to control remote robots and is essential for collecting real-world motion data. Approaches like the Teleoperated Whole‑Body Imitation System (TWIST), recently published in May 2025 (arXiv:2505.02833), extends this by combining motion‑capture retargeting and a unified Reinforcemen Learning (RL)+Behavior Cloning (BC) controller to achieve versatile, coordinated whole‑body robot skills including manipulation, locomotion, and expressive motion. A training dataset is constructed by retargeting large-scale MoCap datasets (e.g. AMASS, OMOMO) to humanoid robots, and then a single controller is trained via RL+BC in simulation, which we expect to transfer to our Unitree G1.
Research Question
How well does this pipeline—retargeting, RL+BC controller training, and real‑world deployment—perform on our humanoid platform using our own MoCap dataset and robot hardware? Specifically:
- How accurate and robust is the tracking performance?
- What differences emerge compared to published TWIST results?
- How sensitive is performance to aspects such as motion diversity, delay, and force-related tasks?
Thesis Goals & Methodology
- Develop a motion‑retargeting pipeline using our human motion capture data.
- Train a unified RL+BC whole‑body controller in simulation.
- Deploy the controller on our Unitree G1 robot (excluding finger movements) and evaluate whole‑body capabilities including locomotion and expressive motions.
- Perform comparative analysis metrics: tracking accuracy, robustness, generalization to novel motions, physical locomotion tasks, and latency.
- Analyze key factors influencing performance: size and diversity of dataset, inclusion of privileged future motion frames, in‑house MoCap noise etc.
Your profile
- Firm understanding of machine learning techniques and reinforcement learning in particular
- At least intermediate Python programming skills
- Experience with ROS2 robotics middleware, simulation and/or humanoids is a plus
About us
We are a high performance team with former industrial experience from Google and McKinsey and are searching for highly motivated students eager to learn and to shape the future of Embodied AI. Our project is funded by the prestigious EXIST Forschungstransfer grant and aims at incorporation.
Available Resources
We provide access to our motion‑capture suit, compute infrastructure for simulation and RL training, and a Unitree G1 humanoid robot (excluding finger movements).
How to Apply
Send an email to julian.ilg@tum.de with the following attachments:
- Your CV
- A brief statement of motivation, including relevant background in robotics, RL, control or motion capture
- Your preferred start date
- Academic transcripts
We look forward to your application!
References
- Ze et al. (2025). TWIST: Teleoperated Whole‑Body Imitation System. arXiv:2505.02833 [cs.RO]
Kontakt: julian.ilg@tum.de