RoboCup 2023 Humanoid AdultSize Winner NimbRo: NimbRoNet3 Visual Perception and Responsive Gait with Waveform In-walk Kicks

TL;DR

The paper addresses the challenge of robust perception, stable walking, and strong kicking for humanoid soccer robots in RoboCup. It introduces NimbRoNet3 for real-time opponent pose estimation and field understanding, augments balance with extended fused-angle feedback and a COM-ZMP controller, and enables stronger, online-tunable kicks via waveform-based in-walk kicking. A five-mass gait for push recovery and a unified control framework culminate in superior performance, with the team winning the RoboCup 2023 AdultSize Soccer Tournament and achieving a flawless defensive record. The contributions demonstrate practical, real-time perception and control improvements that enhance resilience to disturbances and game pace in humanoid soccer robots.

Abstract

The RoboCup Humanoid League holds annual soccer robot world championships towards the long-term objective of winning against the FIFA world champions by 2050. The participating teams continuously improve their systems. This paper presents the upgrades to our humanoid soccer system, leading our team NimbRo to win the Soccer Tournament in the Humanoid AdultSize League at RoboCup 2023 in Bordeaux, France. The mentioned upgrades consist of: an updated model architecture for visual perception, extended fused angles feedback mechanisms and an additional COM-ZMP controller for walking robustness, and parametric in-walk kicks through waveforms.

Figures (7)

• Figure 1: RoboCup 2023 in Bordeaux. Left: Team NimbRo AdultSize with NimbRo-OP2(X) robots, Right: Scene from the Final soccer game vs. HERoEHS (Korea).
• Figure 2: NimbRoNet3 model. Our model employs an encoder-decoder architecture with a pretrained ResNet-18 backbone, and four different network heads for field segmentation, object detection, and HR/LR robot pose estimation.
• Figure 3: Qualitative results for NimbRoNet3. a) Captured images. b) Predicted object heatmaps for soccer balls (red), robots (blue) and goalposts (green). c) Semantic segmentation of the field with lines (white), field (gray) and background (black). d) Predicted robot poses. We depict the predicted joint locations with a circle and connect all joints corresponding to a single robot.
• Figure 4: In-walk kick waveforms. Left: Isolated design of the retraction waveform. Middle: Isolated design of the swing waveform. Right: Absolute applied values with scaling to the abstract gait pose. $\phi_k$ is the kick phase (like $x_K$ in Fig. 6 of nimbro_winners_2019). $s_{swing}$ and $s_{retract}$ are the experimentally achieved swing and retract waveforms, the gray area depicts the time-window where momentum is the highest, e.g. leg is being fully extended and swung forward.
• Figure 5: Technical Challenge: Push Recovery. The robot successfully recovers from a frontal push of 10kg pendulum, which is more than half of the robot's weight.