Bit-Bots Extended Abstract 2025

Jan Gutsche, Jasper Guldenstein

Universität Hamburg, Hamburg, Germany

Abstract At RoboCup 2023, the team demonstrated general functionality of their robot system but identified major reliability issues with hardware and software components. By concentrating on steady improvements and reliability, they established a strong foundation. For RoboCup 2024, they further enhanced maintenance procedures, game preparation processes, and implemented live monitoring tools that proved vital during competition. Looking forward to RoboCup 2025, the team addresses remaining hardware reliability challenges, team member acquisition and training due to student graduation cycles, and improvements to object detection in edge cases. Major planned changes include replacing their Wolfgang Robot platform with a new BitBan platform based on Rhoban's Sigmaban, upgrading their vision pipeline from YOEO to YOLOv7, implementing proper throw-in motions, transitioning path planning to a visibility graph approach in Rust, and exploring zero-shot transfer reinforcement learning policies for robot motion from simulation.

Lessons learned in previous RoboCup competitions

At RoboCup 2023, we showed the general functionality of our robot system. However, we noticed major issues with the reliability of some soft- and hardware components. Therefore, we concentrated on steady improvements without relying on large algorithmic changes. Once a satisfactory level had been reached, we were ready to experiment with new ideas and even introduce and validate new features during the competition. Focusing on reliability also meant improving our maintenance procedures: We regularly and thoroughly checked the robot's hardware and replaced parts that we previously deemed acceptable.

Our efforts to simplify game preparations and adhere to our processes helped reduce errors and stress immediately before a competition; this included tools to enable software deployment to all robots by one person simultaneously and automatic verification of correct software versions. Additionally, we assigned fixed roles to our members (deployment, monitoring, robot handler, note-keeper, repair, recording) and assigned protocols to quickly make an informed decision and execute. We developed live monitoring tools, which proved vital during the competition. Monitoring enabled us to better understand what went wrong during failures, as we had detailed notes and recordings. Using this knowledge, we could find and fix problems more quickly.

All these changes showed worthwhile at the 2024 competition.

Major problems the team is trying to solve for RoboCup 2025

During the 2024 competition, we faced some hardware problems. While we could reduce these compared to the previous competition, much time was still invested in maintenance.

As the Hamburg Bit-Bots are a student team, there is a fluctuation in members as they graduate. We focused our efforts this year on acquiring and introducing new team members to ensure future participation in the competition.

Our vision pipeline performed exceptionally well in previous years in the context of the robot's overall functionality. This year, we noticed that we can improve object detection in edge cases (especially for very close or far away objects).

Plans for the major changes the team expects to have implemented for RoboCup 2025

We are replacing our Wolfgang Robot platform with a new platform called BitBan based on team Rhoban's Sigmaban platform [1]. It will feature the same kinematic structure as the Sigmaban, but we will use our custom solutions for electronics. We have mostly finished development and prototyping and are now manufacturing and assembling the robots.

Our particle filter-based localization platform performed reasonably well in previous competitions. Due to processing time in the vision pipeline, updates are calculated on delayed measurements. While we already compensated for this in the inverse perspective mapping, we are now also accounting for it in the particle update. We have implemented this change and are currently testing its robustness.

We are upgrading our vision pipeline YOEO [3] to be based upon the more modern YOLOv7 [4] architecture. We have prepared previously recorded and additional training data and started the implementation.

Furthermore, we already implemented a proper throw-in motion and behavior that will be used instead of the previously kick-in.

In addition to that, we are in the process of moving our path planning from a grid based to a visibility graph based approach implemented in Rust.

Recently, the paradigm of zero-shot transfer RL policies for robot motion from simulation has emerged [2]. We are currently developing simulation environments and reward functions to enable this for our robot platform. We plan to have a working prototype, which, depending on performance, may be employed in the 2025 competition. As this is an experimental approach, we are also investing some effort in the continued development of our current walking algorithm.

References

1. Julien Allali et al. "Rhoban Football Club: RoboCup Humanoid Kid-Size 2023 Champion Team Paper". In: RoboCup 2023: Robot World Cup XXVI. Ed. by Cédric Buche et al. Springer Nature Switzerland, 2024, pp. 325–336.
2. Tuomas Haarnoja et al. "Learning agile soccer skills for a bipedal robot with deep reinforcement learning". In: Science Robotics 9.89 (2024).
3. Florian Vahl et al. "YOEO -You Only Encode Once: A CNN for Embedded Object Detection and Semantic Segmentation". In: IEEE International Conference on Robotics and Biomimetics (ROBIO). Dec. 2021.
4. Chien-Yao Wang, Alexey Bochkovskiy, and Hong-Yuan Mark Liao. "YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors". In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). June 2023, pp. 7464–7475.