Careers and Jobs

Background

Assembly is a crucial skill for robots in both modern manufacturing and service robotics. However, mastering transferable insertion skills that can handle a variety of high-precision assembly tasks remains a significant challenge [1][2]. In our previous work [3], we leveraged the generalization capacity of the diffusion model and adapted it to force-domain tight-clearance assembly tasks, achieving remarkable transferability to novel scenarios.

In this work, we aim to explore the possibility to use flow matching [4] (a recent framework for generative modeling that has achieved state-of-the-art performance across various domains) to imitation the successful demonstrations collected from previous reinforcement learning training.

Your Task

• Understand our previous solution(code) for and tight-clearance industrial insertion tasks with for force domain wiggle motion [1,2,3].
• Learn the knowledge regarding flow matching and read related papers in the field of robot manipulation.
• Train a flow matching based imitation learning algorithm, and later on implement it on the real-world robot.
• Conducting experiments and analyze the results.

Requirement

• Strong self-motivation and ability to work independently.
• Background in robotics, with relevant coursework or practical experience.
• Proficiency in Python, with hands-on experience in PyTorch.
• Familiarity with diffusion models and flow matching techniques is a plus.

Supervisor: Yansong Wu, Dr. Zewen Yang

To apply, send your CV, transcript, and supporting documents to yansong.wu(at)tum.de and zewen.yang@tum.de before 31 May 2025.

TUM School of Computation, Information and Technology 

Technische Universität München

Reference:

1. Johannsmeier L, Gerchow M, Haddadin S. A framework for robot manipulation: Skill formalism, meta learning and adaptive control[C]//2019 International Conference on Robotics and Automation (ICRA). IEEE, 2019: 5844-5850.
2. Wu Y, Wu F, Chen L, et al. 1 khz behavior tree for self-adaptable tactile insertion[C]//2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024: 16002-16008.
3. Wu Y, Chen Z, Wu F, et al. TacDiffusion: Force-domain Diffusion Policy for Precise Tactile Manipulation[J]. arXiv preprint arXiv:2409.11047, 2024.
4. Dai, Xiaobing, Dian Yu, Shanshan Zhang, and Zewen Yang. "Safe Flow Matching: Robot Motion Planning with Control Barrier Functions." arXiv preprint arXiv:2504.08661 (2025).

Background:

Shared autonomy refers to a collaborative control paradigm wherein both a human operator and an autonomous robotic system share the responsibility of executing a task [1]. This approach leverages the strengths of human intelligence—such as intuition, adaptability, and decision-making—alongside the precision, repeatability, and computational power of robots. Consequently, this system is capable of managing complex tasks while enhancing efficiency, safety, and usability.

In this study, we focus on addressing the industrial assembly task through teleoperation skills. However, due to low transparency of the system and human-interaction manner, tackling contact-rich manipulation tasks, particularly those involving tight-clearance manipulation, remains a significant challenge. To remedy this gap, we propose integrating the knowledge gained from robotic assembly tasks into teleoperation within a shared-autonomy framework.  

Your Tasks:

1. Understand our previous solution(code) for solving the shared autonomy teleoperation work [2] and tight-clearance industrial insertion tasks with for force domain wiggle motion [3,4].
2. Propose the autonomy allocation method in our application.
3. Integrate the force domain wiggle motion into our teleoperation system based on the shared autonomy under our guidance.
4. Make experiments to demonstrate the feasibility and superiority of this method.

Requirement:

• Highly self-motivated;
• Experiences or knowledge from related Robotics courses;
• C++ and python programming experience.

To apply:

Send your personal CV and transcript as attachment to both yansong.wu(at)tum.de and xiaoyu.chen(at)tum.de.
 

Job Description:

t3://file?uid=4525

Recent advancements in robotics, especially concerning humanoids and quadrupeds, are largely due to the adoption of novel actuator technologies. These technologies are involving high power density BLDC motors with a lower gear ratio, with the aim of maintaining good proprioceptive feedback for control. [1] However, active development of new actuation concepts is ongoing. One of the important directions is an augmentation of such actuators with mechanical springs for storing and releasing energy at the dynamic peaks.

Our work focuses on the development and testing of one such actuator, a version of the Parallel Elastic Actuator. This particular task includes the manufacturing of a simple test-actuator that integrates a backdrivable motor with a spring and a torque sensor connected in parallel. Further, a performance characterization of the motor will be conducted (Bode plot analysis, etc.). Additionally, an impedance controller will be evaluated on such a setup.

What you will gain:

• Experience in Modeling and Control of Robotics systems
• Experience building, prototyping
• Best design practices for torque sensor integration in actuators
• Insights into our System Development and access to our community
• BLDC motor control

Requirements from candidates:

• Mechanical Engineering background
• Completed classical control courses (or some project experience in control)
• Any CAD software for the Part designs (such as Solidworks, Fusion 360,etc.)
• Matlab skills
• Basic skills in Electronics
• Plus are:
  • Understanding how Motors work
  • Familiarity with GIT
  • Working skills in Ubuntu operating system

To apply, you can send your CV, and short motivation to:

Supervisors 

[1] P. M. Wensing, A. Wang, S. Seok, D. Otten, J. Lang and S. Kim, "Proprioceptive Actuator Design in the MIT Cheetah: Impact Mitigation and High-Bandwidth Physical Interaction for Dynamic Legged Robots," in IEEE Transactions on Robotics, vol. 33, no. 3, pp. 509-522, June 2017, doi: 10.1109/TRO.2016.2640183.