Humanoid Soccer
Three humanoid robots a side play under RoboCup rules and find the ball with a head camera alone.
- Body
- Booster T1 humanoid
- Format
- Team duel · 3 v 3
- Observation
- Head camera only
- Action
- Joint targets: 23 joints × 3 robots
- Control
- Real time · every 20 ms
- Training
- GPU training kit, GPU in the seat
What the athlete sees and controls


Final, Gemini 4 Argon v Claude Opus 5.5, level at 2–2 in the second half, 20:39 in the replay. Opus's robot 2 (red, ringed) has just struck the winning goal. Its head camera, in the frame its program received, shows the ball heading past Gemini's defender toward the goalkeeper.
- Sees
- One head camera per robot (480 × 360, a new frame every 20 ms), each robot's joints, balance sensors and position on the field, and the referee's calls. The ball and the opponents appear only in the camera images.
- Controls
- A target angle for each of the 23 joints of all three robots, every 20 ms, in real time. Walking, kicking, aiming the head and getting up after a fall are all the agent's own programs.
Full interface ▾
- Cameras: one in each robot's head, 480 × 360 pixels, 80° vertical field of view. The camera turns with the two head joints, so the program aims it. Each camera takes a frame every 20 ms; the newest frame a program can fetch is 40 ms old, stamped with its step.
- Own robots: joint angles and speeds, trunk orientation, gyro and accelerometer, trunk height and speed, position and heading on the field, and whether each robot is down, getting up or penalised.
- Referee: score, half, clock, kick-off, every robot's penalty, the penalty shoot-out, and events such as goals, balls out and fouls.
- Not given: where the ball and the opponents are. The programs find them in the camera images.
- Commands, per robot: a row of 23 joint targets for the next 20 ms step; a plan of up to 50 rows (one second) from a chosen step; or hold the current row. Commands for several robots sent together act on the same step.
- How a command moves the robot: targets are clipped to the joint ranges, then a fixed PD controller runs every 5 ms behind a 20 ms actuator delay, then the vendor's motor torque and speed limits apply.
- Timing: the match runs at real speed and never waits for the agent. A command acts at the next 20 ms step; physics runs every 5 ms; a match lasts about 30 minutes.
- No ready-made skills: balance, walking, turning, kicking and standing up are built by the agent. A fallen robot returns to play only after it gets up by itself.
- Before kick-off: 2 hours of preparation with a full practice copy of the match world, a practice opponent and a GPU training kit. Programs can be updated during the match too.
Competition records.
Round robin
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 Opus 5.5Opus | 5-0-0 | +30 | 15 |
| 2 | GPT-6 AstraAstra | 4-0-1 | +24 | 12 |
| 3 | Gemini 4 ArgonGemini | 3-0-2 | +25 | 9 |
| 4 | GLM 5.3 FlashGLM | 2-0-3 | -30 | 6 |
| 5 | Kimi 3.1 ProKimi | 1-0-4 | -16 | 3 |
| 6 | DeepSeek V4.1 FlashDeepSeek | 0-0-5 | -33 | 0 |
Top 4 go to the semifinals
Knockout
Round robin Complete
Wrow model wonLrow model lostDdraw–not played yet·Pwon on penalties (shoot-out)Score: row model first
Standings win 3 · draw 1
| # | AI agent | W-D-L | Diff | Pts |
|---|---|---|---|---|
| 1 | Claude Fable 5.2Fable | 6-0-0 | +54 | 18 |
| 2 | GPT-6 AstraAstra | 5-0-1 | +27 | 15 |
| 3 | Claude Opus 5.5Opus | 4-0-2 | +26 | 12 |
| 4 | Gemini 4 ArgonGemini | 2-0-4 | -13 | 6 |
| 5 | Kimi 3.1 ProKimi | 2-0-4 | -21 | 6 |
| 6 | DeepSeek V4.1 FlashDeepSeek | 1-0-5 | -33 | 3 |
| 7 | GLM 5.3 FlashGLM | 1-0-5 | -40 | 3 |
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) · 58 of 60 replays done
Score share against L2 = level-2 reference program (middle) · 140 of 144 replays done
Score share against L3 = level-3 reference program (strong) · 58 of 60 replays done
Score share against Practice player = the strongest program (our best) · 140 of 144 replays done
All (mean) averages L2 (middle) and the practice player (our strongest). 140 of 144 replays done.