A Hierarchical, Model-Based System for High-Performance Humanoid Soccer

TL;DR

The paper addresses the challenge of robust, fully autonomous adult-sized humanoid soccer by delivering ARTEMIS, a tightly integrated hardware–software stack that combines high-performance actuation, a specialized foot design, and a perception–localization–planning–control framework. It introduces a model-based control philosophy with dynamic planning (DAVG), collision-aware cf-MPC, CLAP localization, and Perception-Locked Mid-Swing Kicking (PLMK) to achieve fast, stable, in-gait kicking and agile navigation. Through extensive in-field testing, high-fidelity simulations, and RoboCup 2024 results, the work demonstrates reliable perception, precise localization, real-time obstacle avoidance, and effective cooperative play, culminating in a RoboCup Adult-Sized Humanoid Soccer champion. The discussion highlights the value of tight integration and identifies pathways toward hybridizing model-based approaches with learning-based adaptability for even greater autonomy in dense, adversarial environments.

Abstract

The development of athletic humanoid robots has gained significant attention as advances in actuation, sensing, and control enable increasingly dynamic, real-world capabilities. RoboCup, an international competition of fully autonomous humanoid robots, provides a uniquely challenging benchmark for such systems, culminating in the long-term goal of competing against human soccer players by 2050. This paper presents the hardware and software innovations underlying our team's victory in the RoboCup 2024 Adult-Sized Humanoid Soccer Competition. On the hardware side, we introduce an adult-sized humanoid platform built with lightweight structural components, high-torque quasi-direct-drive actuators, and a specialized foot design that enables powerful in-gait kicks while preserving locomotion robustness. On the software side, we develop an integrated perception and localization framework that combines stereo vision, object detection, and landmark-based fusion to provide reliable estimates of the ball, goals, teammates, and opponents. A mid-level navigation stack then generates collision-aware, dynamically feasible trajectories, while a centralized behavior manager coordinates high-level decision making, role selection, and kick execution based on the evolving game state. The seamless integration of these subsystems results in fast, precise, and tactically effective gameplay, enabling robust performance under the dynamic and adversarial conditions of real matches. This paper presents the design principles, system architecture, and experimental results that contributed to ARTEMIS's success as the 2024 Adult-Sized Humanoid Soccer champion.

Figures (12)

• Figure 1: Overview of the ARTEMIS humanoid soccer system. A). Two ARTEMIS humanoid robots competing for ball possession during a practice match. B). Perception and localization pipeline: stereo vision detects field landmarks, the ball, teammates, and opponents; CLAP fuses landmark geometry with inertial measurements to estimate the robot pose ($x_{r}$, $y_{r}$, $\theta_{r}$). Proximity sensing provides complementary obstacle information when vision is degraded. C). Mid-level navigation and behavior framework: the robot selects a desired pose and shot target based on the game state, and a collision-aware MPC controller generates a dynamically feasible trajectory that respects proximity constraints and field geometry.
• Figure 2: System architecture of the ARTEMIS humanoid platform. The perception layer provides object detections, proximity cues, and pose estimates that feed into the path planner and behavioral planner module. A high-rate shared-memory loop (SHM) handles hardware interaction and executes the locomotion controller.
• Figure 3: Foot Design a). Overview of three versions of design. b). Isometric view and sectional view of bolt version. c). Isometric view and sectional view of magnet version. d). Isometric view and sectional view of buckle version. e). Top view of selected (buckle) version assembly. f). Improvement of support structure for Front/Side Bill
• Figure 4: Visualization of the Geometric Field-of-View approach, filtering out expected depth map readings and appending alien depth map values within the proximity field.
• Figure 5: Localization method. a). Localization framework overview. b). Probability distribution of localization cluster.