WUT students reach the finals of a robotics competition in Shenzhen

The Humanoid Student Research Group operates at the Faculty of Mechatronics of the Warsaw University of Technology, with its team also including students from the Faculty of Electronics and Information Technology. The project is carried out by Arkadiusz Cybulski, Piotr Gerke, Konrad Suchodolski, Nikodem Kuliś, Walery Kusznirski, Jacek Caban, and Piotr Zarzycki.

What can the robot do?

The competition task was to train a humanoid robot to perform dynamic combat scenarios, including boxing sequences and movements requiring it to maintain balance during fast, abrupt manoeuvres. Instead of programming every movement manually, the team used reinforcement learning: the robot learns to move through trial and error in a virtual environment, performing millions of repetitions in simulation before its newly acquired skills are transferred to the physical robot. The starting point for the training was the movements of a real human.

“We developed our own low-cost motion capture system using commercially available equipment—Kinect cameras from the Xbox console,” says Konrad Suchodolski. “The recorded movements then had to be transferred to the robot’s body in a process known as retargeting, because the robot has different proportions, weight, and joint limitations than a human. The collected and processed motion data were added to a shared motion library, which was used by an automated training pipeline running on graphics processing units (GPUs). This automation allowed us to test successive variants quickly and compare the results, instead of configuring each training run individually.”

The project currently focuses on two main pillars: artificial intelligence models for motion control and proprietary software solutions that incorporate the robot’s mechanical constraints and specifications directly into the virtual environment.

The greatest challenge

Since all work at this stage was carried out exclusively in simulation, the team's biggest challenge at this point was not transferring the algorithms to a physical robot, but rather developing an efficient workflow and achieving high-quality motion retargeting.

The need to manage vast amounts of data, automate the training process, and optimise computation time required the team to design an efficient training pipeline from the ground up.

“We had to develop an architecture that allowed us to iterate quickly and test multiple artificial intelligence models in a short period of time,” says Konrad Suchodolski.

Transferring movements recorded from a human performer to a robot model with entirely different kinematics, proportions, and weight distribution also proved to be a complex task. The movements in the virtual environment had to appear natural, preserve the dynamics of the strikes, and at the same time strictly comply with the robot’s physical constraints so that the AI model could learn from them correctly.

The Brave Mind Fighters team

What’s next for the Humanoid Student Research Group?

The next stages of the competition will take place in Shenzhen. There, students from the Warsaw University of Technology will deploy and test their developed models directly on physical robots provided by EngineAI. It will be their first opportunity to evaluate their solutions on the target hardware under competition conditions and to tackle the so-called sim-to-real gap—the difference between simulation and the real world.

The solutions developed by the students, as well as the humanoid project itself, have attracted the attention of not only the scientific community but also the general public. The robot has generated considerable interest from national media, and the team’s preparations for the international finals have recently been featured in television programmes, further promoting the scientific achievements of the Warsaw University of Technology.