Humanoid soccer stops juggling separate controllers: SoccerBot unifies four skills on a human motion prior and transfers zero-shot to a real robot
Synopsis
Humanoid soccer robots have either covered only a few skills or skipped motion priors entirely and moved unnaturally. SoccerBot distills large-scale human soccer motion-capture data into a structured latent motion prior, letting a single skill-conditioned policy learn moving, dribbling, trapping and kicking in one latent space. On the Unitree G1, the policy transfers zero-shot to hardware, with simulation success rates of 97.04% for kicking and 89.88% for dribbling, while real-robot kicking direction error rises from 4.09 to 10.41 degrees.
Interpretation
To the authors' knowledge, this is the first real-world humanoid soccer system that combines a complete soccer skill repertoire with human-like motion quality, with moving, dribbling, trapping and kicking handled by a single policy. Earlier soccer systems mostly handled one or two of these skills, and multi-skill systems relied on separately trained controllers switched by an explicit state machine, producing discontinuous transitions; multi-skill soccer with human motion priors had been shown only for simulated characters. In simulation episodes per skill, the policy reaches 97.04% kicking and 89.88% dribbling success, above the compared PPO, AMP, ASE and LATENT methods; it completes every moving command and traps 83.08% of passes.
The policy does not output joint targets directly; it learns a residual inside a frozen latent motion space, so exploration stays near human motion. Task-reward-only reinforcement learning without a motion prior can kick and trap but never dribbles successfully, and its motion lacks naturalness. Against human reference motions, the policy's MMD2 distance is 0.226 on the training reference set and 0.554 on motion clips excluded from training, the closest among methods that complete all four skills.
Training initializes each skill from the terminal states of a preceding skill, so the policy sees the posture, momentum and ball configuration left by the switch. From canonical starts, having or omitting this initialization differs by only a few points on kicking and dribbling. Starting from other skills' terminal states, the full model is higher in most non-trapping switching cells, such as 87.5% versus 78.0% for dribble-to-kick and 78.5% versus 71.5% for move-to-dribble; across four ball and command conditions the full model's kicking success exceeds the variant without this initialization in every condition, with the largest gap on mid-range balls at 87.0% versus 34.3%.
Perspective
The demonstrated scope is four basic soccer skills on the Unitree G1: episodes per skill in simulation and trials per skill on hardware, with ball and robot states supplied by an external optical motion-capture system and transformed into the robot base frame.
The sim-to-real gap varies by skill: on hardware kicks overshoot the commanded speed, and dribbling and trapping show the largest increases in continuous error metrics, which the authors suggest may partly stem from the simulated ball being rigid while the real inflated ball deforms on impact. The system relies on external motion capture, restricting it to instrumented indoor settings; the skill set is limited to four basic behaviors, and experiments cover single-robot control only.