Source-linked AI summary
B-Human '26 - Scalable Team Play up to 11 vs. 11
Arne Hasselbring, Tim Laue, Thomas Röfer, Philip Reichenberg
TL;DR
The new HSL’s varied field layouts and player counts called for adapting B-Human’s system to the new platform and rules. The paper describes a behavior approach informed by other RoboCup leagues, and reports that B-Human secured every 2026 soccer event it entered.
Problem
The HSL introduced multiple field layouts and player counts, requiring adaptation to the new rules and configurations.
Source: §1 · p. 1Method
B-Human describes a humanoid team behavior approach that adopts and adapts concepts from RoboCup’s Simulation and Small-Size Leagues.
Source: §1 · p. 2Results
In 2026, B-Human secured every soccer event it entered, with the highest goals per minute and fewest penalties per game among participating teams.
Source: §4 · pp. 10–11Takeaways & Limitations
The behavior approach won B-Human’s 3 vs. 3 and 5 vs. 5 Middle Division games and scaled toward 11 vs. 11 in the first humanoid matches.
Source: §5 · p. 11Future work includes more sophisticated 11 vs. 11 behaviors and, first and foremost, tracking and playing flying balls.
Source: §5 · p. 12
Abstract
from arXiv · showhide
B-Human won multiple competitions in the Middle Division of the new Humanoid Soccer League in 2026. In addition to regular matches with three or five players per team, we ventured into uncharted territory by playing demo matches with eleven humanoid robots, highlighting the tactical flexibility of our behavior and the overall robustness of our system. As B-Human formerly competed in the Standard Platform League with robots half the size and less than a third the weight, we begin by outlining the adjustments needed for this transition. The central section of this paper details our behavior architecture, drawing on concepts from RoboCup leagues that have used 11 players for much longer.
1 Introduction
B-Human adapted to the new Humanoid Soccer League and its varied playing configurations after moving to the Booster K1 platform. The paper’s main contribution is its humanoid behavior approach, informed by RoboCup leagues with established multi-agent play.
1 Introduction
- After the Standard Platform League and Humanoid League merged into the HSL, B-Human chose the Booster K1 for the Middle Division, building on prior work with the Booster T1.Source: §1 · p. 1
- The new rules introduced varied fields and team sizes, including the first humanoid 11 vs. 11 match; B-Human’s system is intended to support teams from 1 to 11.Source: §1 · pp. 1–2
- The paper describes a humanoid behavior approach adapted from the Simulation and Small-Size Leagues, preceded by a system overview and followed by 2026 competition results.Source: §1 · p. 2
2 System Overview
The system extends B-Human’s NAO software to the Booster K1, retaining some components while replacing others. Its overview covers object detection, state estimation, and reinforcement-learned motion skills.
2 System Overview
- The system retains or expands its modular NAO design of well over 100 components, keeping some with minor changes while replacing detection and motion solutions.Source: §2 · p. 2
2.1 Object Detection
- YOLOX detects balls and robots at 60 Hz; geometric refinement addresses wide-angle distortion, while retained line detection helps reduce localization error.Source: §2.1 · pp. 3–4
2.2 State Estimation
- Localization uses a small-particle filter with an Unscented Kalman filter per particle; goals are excluded, and accuracy lacks systematic external evaluation.Informal inspection indicates typical deviation below approximately 10 cm.Source: §2.2 · p. 4
- Ball tracking maintains stationary and rolling hypotheses and predicts rolling trajectories; teammate models fuse messages with observations, while opponents are limited to recently seen robots.Source: §2.2 · p. 5
2.3 Motion Skills
- Reinforcement learning was used to learn walking, kicks of different strengths, and standing up after a fall.Source: §2.3 · p. 5
3 Behavior
B-Human’s behavior combines team-level tactics and position assignment with individual positioning and passing decisions. Its architecture draws on decentralized coordination and supports teams with different player counts.
3 Behavior
- Behavior pairs configuration-based team coordination with CABSL2 individual skills, using wireless LAN without reverting to overly centralized coordination.The HSL allows 12,000 messages per team per match, and the architecture benefits from up-to-date teammate data.Source: §3 · p. 5
3.1 Strategy
- Strategies govern available tactics, transitions, and game-state set plays; tactics encode labeled positions and priorities, while set plays shift selected locations and tactics or free kicks use team-message majority votes.Set plays also restrict the region where the ball may be, and votes are evaluated locally with ties broken by the lower-numbered player’s choice.Source: §3.1 · p. 6
3.2 Position / Role Assignment
- Players independently hypothesize a team-wide assignment using communicated locations, prior-assignment discounts, and tactic or set-play mirroring when it lowers the overall cost.The goalkeeper is assigned by player number, while the remaining field-player assignments are costed by distance to target positions.Source: §3.2 · p. 7
- With ten field players assigned, the ranked-cost Hungarian method is roughly 7000 times faster than brute force and runs well below 1 ms.It maps ranked assignment costs to integers represented by single bits, allowing the Hungarian method to solve the stricter bottleneck assignment problem.Source: §3.2 · p. 8
- Active roles temporarily override position roles, while a ball-chasing player retains its tactical place, generally spacing teammates out for passing.A player may also take the closest-to-team-ball role to search for the ball when it is considerably closer than the current ball player.Source: §3.2 · p. 8
3.3 Positioning
- Players optimize role-specific ratings within precomputed Voronoi cells, with role-dependent movement preferences and safeguards against jittering or unnecessary walking.A player starts walking only when the rating at its target is considerably better than at its current position.Source: §3.3 · pp. 9–10
3.4 Passing
- Passes can outrank dribbling or direct shots; a teammate is selected using combined passing and scoring probabilities with hysteresis, only when direct scoring is less likely.A stronger kick can reduce interception risk by shortening the time the ball rolls outside the team’s control.Source: Figure 7 · p. 10, §3.4 · p. 10
4 Results
B-Human secured every soccer event it entered in 2026, leading participating teams in goals per minute and penalties per game. The reported outcomes reflect inherited team-play software combined with new walking and kicking skills.
4 Results
- 1.24 goals per minute and 0.4 penalties per game at the RoboCup were the highest and fewest, respectively, among participating teams.Table 1 presents B-Human’s 2026 HSL Middle Division game results.Source: §4 · pp. 10–11, Table 1 · p. 11
- Existing NAO software with complex team play and passing scaled across setups, while new walk and kick skills made it effective on the new robots.Source: §4 · p. 11
- Rapid, accurate kicks with varied strengths support dribbling, passing, and shots from any position; the walk is described as fast and robust.Source: §4 · p. 11
5 Conclusion and Future Work
The paper describes B-Human’s behavior approach for the new humanoid platform, building on solutions from the Standard Platform League. The team met its 2026 objectives, won all Middle Division 3-v-3 and 5-v-5 games, and demonstrated behavior scaling to 11-v-11.
5 Conclusion and Future Work
- The comprehensive team behavior approach won all Middle Division games in 3-v-3 and 5-v-5 matches and scaled toward 11-v-11.The latter capability was demonstrated in the world’s first humanoid 11-v-11 matches.Source: §5 · p. 11
- With less than six months to work with the robots, B-Human met all key objectives of getting the system running and addressing issues needed to perform well.Source: §5 · p. 11
- Future work includes more sophisticated 11-v-11 behaviors and, first and foremost, tracking and playing flying balls.Source: §5 · p. 12