Wheeled Soccer
Eleven fast wheeled robots a side, RoboCup Small Size League rules, and a strongest opponent built on a RoboCup champion team's open-source software.
- Body
- Omnidirectional small-size wheeled robot
- Format
- Team duel · 11 v 11
- Observation
- Shared overhead vision data and referee messages
- Action
- Velocity, kick and dribbler for each of 11 robots
- Control
- Real time · every 10 ms
- Training
- GPU training kit, runs on the CPU (no GPU in the seat)
What the athlete sees and controls


Semifinal 2, GPT-6 Astra v Gemini 4 Argon (3–3, Astra through on penalties), 10:26 in the replay. Astra's shot (blue) is on its way past Gemini's defenders; half a second later it is 3–2. The Final had only own goals, so this card uses the semifinal.
- Sees
- No camera: the league's vision system, as real SSL teams get it. About 67 times a second, 40 ms late and noisy: every visible robot's number, team, position and heading, and the ball's position, plus the referee's tracker and game-controller messages.
- Controls
- A radio command for each of the 11 robots every 10 ms, in real time: drive velocity forward, sideways and turning, kick speed (flat or chip) and dribbler speed, the same command real SSL software sends.
Full interface ▾
- No camera images. The robots see the game the way real RoboCup Small Size League teams do: through the league's overhead vision system, delivered as numbers.
- Vision: one packet per overhead camera and frame, about 67 Hz, 40 ms after capture, with realistic noise, missed and false detections. Each packet lists every robot it sees (number, team, x, y, heading) and the ball (x, y).
- Own robots appear only in the vision data. The match gives no position or speed of them, only their radio feedback: ball in the break beam, kicker charge, dribbler and armed kick.
- Referee: the automatic referee's tracker (filtered positions and speeds) and the game controller's messages (state, score, cards, ball placement, free kicks).
- Command per robot: driving speed forward and sideways and turning rate in the robot's own frame, kick speed and angle (0 flat, above 0 chip), dribbler speed. Each command replaces the last one; a field left out is off.
- Robot limits (after the TIGERs Mannheim 2020 robot): 3 m/s, 4.7 m/s² acceleration, 5.9 m/s² braking, 20 rad/s turning; flat kick up to 8.5 m/s, chip up to 5.5 m/s; dribbler up to 17000 rpm. A robot whose command is not renewed within 0.1 s brakes to a stop.
- The kicker is armed while a command with a kick speed is current and fires the moment the ball is in the beam. Substitutions are their own request, needed after a yellow or red card.
- Timing: a control step every 10 ms, physics every 1 ms; the match never waits. Two halves of 5 minutes of playing time; with ball placements and free kicks a match takes up to about 2 hours.
- Rules: RoboCup SSL Division A, 11 v 11 on a 12 × 9 m field; ball placement after a stop is the team's own job.
- Before kick-off: 60 minutes of preparation with a full practice copy of the match world and the TIGERs Mannheim software as practice opponent. No GPU.
Competition records.
Round robin
Wrow model wonLrow model lostDdraw–not played yet·Pwon on penalties (shoot-out)·Tdecided by coin toss or lotScore: row model first
Standings win 3 · draw 1
| # | AI agent | W-D-L | Diff | Pts |
|---|---|---|---|---|
| 1 | Claude Opus 5.5Opus | 5-0-0 | +26 | 15 |
| 2 | GPT-6 AstraAstra | 4-0-1 | +23 | 12 |
| 3 | Gemini 4 ArgonGemini | 3-0-2 | +25 | 9 |
| 4 | Kimi 3.1 ProKimi | 2-0-3 | -11 | 6 |
| 5 | DeepSeek V4.1 FlashDeepSeek | 1-0-4 | -23 | 3 |
| 6 | GLM 5.3 FlashGLM | 0-0-5 | -40 | 0 |
Top 4 go to the semifinals
Knockout
Level; penalty shoot-out 3:1
Round robin Complete
Wrow model wonLrow model lostDdraw–not played yet·Pwon on penalties (shoot-out)·Fforfeit: one side did not startScore: row model first
Standings win 3 · draw 1
| # | AI agent | W-D-L | Diff | Pts |
|---|---|---|---|---|
| 1 | Claude Fable 5.2Fable | 6-0-0 | +38 | 18 |
| 2 | Claude Opus 5.5Opus | 5-0-1 | +24 | 15 |
| 3 | Gemini 4 ArgonGemini | 4-0-2 | +27 | 12 |
| 4 | GPT-6 AstraAstra | 3-0-3 | +6 | 9 |
| 5 | Kimi 3.1 ProKimi | 2-0-4 | -19 | 6 |
| 6 | GLM 5.3 FlashGLM | 1-0-5 | -21 | 3 |
| 7 | DeepSeek V4.1 FlashDeepSeek | 0-0-6 | -55 | 0 |
Final standings: Claude Fable 5.2 tops the Episodic Learning board. Episodic Learning: every match starts from an empty workspace.
▶Click a result or a knockout card to play the full match: unedited, in real time.
Checkpoint Evaluation.
Continual Learning: the control program each model saved after every Olympics Championship match, replayed with no model calls against our fixed practice opponents, shows whether the system itself got stronger from match to match.
Score share against L1 = level-1 reference program (weak) · 56 replays
Score share against L2 = level-2 reference program (middle) · 152 replays
Score share against L3 = level-3 reference program (strong) · 56 replays
Score share against Sumatra = TIGERs' SSL program, the practice player (strongest) · 152 replays
All (mean) averages L2 (middle) and Sumatra. 152 of 152 replays done.