To content
Fakultät für Elektrotechnik und Informationstechnik

Project Groups

Development of a digital twin-based reinforcement learning framework for edge-controlled humanoid robots

HERA: Humanoid Edge-controlled Robots with Adaptive Intelligence

-
in
  • Project Group
Group photo of the participants involved in the HERA project group. © ComNets, 2026
Group photo of the participants involved in the HERA project group.
The Master's project group HERA (Humanoid Edge-controlled Robots with Adaptive Intelligence) built an end-to-end pipeline for teaching a humanoid robot new manipulation skills in simulation and transferring them to real hardware. Since training directly on physical humanoids is slow, risky and causes mechanical wear, the eight students first created a digital twin of the open-source PIB humanoid robot in NVIDIA Isaac Lab. There, joint limits, masses, friction, stiffness and damping were matched to measurements of the real robot, so that simulated and physical motion behave alike.

On this basis, the group solved a sequential cube-stacking task with reinforcement learning handled by PPO policies. Training ran across 4096 parallel environments with randomized resets and noisy observations, which noticeably improved robustness and generalization. For execution on the real robot, a stereo camera pipeline locates the cubes including color segmentation, and depth data yielding their 3D positions. The resulting actions are published via ROS 2 sent as joint commands to the physical PIB robot. Finally, the students presented their concept, results and a live demonstration to the members of ComNets Dortmund. The project group is part of the Master's degree program Automation and Robotics and challenges students to solve scientific tasks collaboratively.
 

Further information:

Chair: Chair of Communication Networks (ComNets)
Supervisors: Niklas Wagner, Kevin Šabanović
Start date: October 2025