MESSI  ·  Multi-Expert Soccer Skill Integration

Eight soccer skills.
One controller.

We consolidate eight independently trained soccer experts into a single physics-based policy, then teach it to hand off between skills without forgetting any of them.

Xiaogang Peng1*  ·  Jiaqi Zou2*  ·  Lei Zhong3  ·  Sirui Xu4  ·  Huaizu Jiang1 1Northeastern University    2Duke University    3University of Edinburgh    4University of Illinois at Urbana-Champaign * Equal contribution

At a glance

8

experts

One per skill, each trained with its own motion prior, task reward and curriculum.

1

student

A skill-conditioned transformer, distilled from all eight through one observation space.

+1

adapter

A gated, shared low-rank adapter on the frozen student that learns the handoffs between skills.

The repertoire

Eight skills. One policy.

Off the ball, on the ball, in the air, and back on your feet. Every clip is physics simulation, played back in real time. Click any clip to enlarge it.

Off-ball movement

Runs, cuts and turns to a commanded direction and speed, without the ball.

Dribbling

Keeps the ball close while following a commanded direction and speed.

Stop and turn

Stops the ball under the sole mid-dribble, then turns 180°.

Shooting

Strikes with the commanded foot and technique, at a commanded direction and power.

Passing

Plays ground and lofted passes at a commanded direction and power.

Trapping

Controls an incoming ball on the first touch, with the chest or the foot.

Heading

Meets a ball in the air with the head.

Getting up

Recovers to standing after a fall.

Shooting

Pick a foot. Pick a corner.

Each shot follows a commanded foot, technique, direction and power. The overlay compares the commanded target with the ball's actual speed and direction.

Left foot · Inside · High
Left foot · Inside · Low
Left foot · Instep · High
Left foot · Instep · Low
Right foot · Inside · High
Right foot · Inside · Low
Right foot · Instep · High
Right foot · Instep · Low

Transitions

Same handoff. Different outcome.

Each row starts from one handoff state and runs it three ways, side by side.

  1. Direct switchControl passes to the frozen student at the handover, with no adaptation.
  2. Expert + STI fine-tuneThe target skill's expert, fine-tuned with PPO from states the previous skill left behind, following PhysicsFC's skill-transition initialization.
  3. OursThe frozen student with the shared transition adapter.

Results

Better handoffs. Same skills.

One shared adapter is the most successful on every transition and every chain, and it leaves the native skills exactly where the student had them.

    Paper Tables 2 and 5. Transitions start from identical predecessor-generated states; chains are continuous episodes with no resets, 512 per chain.

    About the baselines
    Direct switch
    The distilled student hands over with no adaptation.
    Expert + STI
    The target expert fine-tuned from predecessor-generated states (skill-transition initialization).
    SkillX*
    One multi-skill policy trained with in-episode skill switches on the same transitions.

    Long horizon

    Control that holds up.

    Long, uncut episodes, played back in real time.

    Dribbling, one uncut episode0:00

    Skill chaining

    One play. No cuts.

    Trap, dribble, stop and turn, dribble, shoot: one uninterrupted episode of our full method. The teammate runs the same policy. It passes, then makes an off-ball run.

    Real time, uncut. The two players are separate episodes joined at the pass; the goal and the net are added at render time.

    Slalom, Shoot!

    Slalom dribbling followed by a shot on goal.

    Pass. Trap. Pass back.

    Two players face to face, both run by the same policy. Each traps the incoming ball, passes it back and steps back to its spot.

    Real time. Two co-simulated episodes share one ball, handed over at each kick;

    How it works

    Integrate the experts.
    Adapt the handoffs.

    MESSI overview. (a) Skill teachers trained with motion priors, task objectives and PPO form a teacher bank. (b) Multi-skill distillation into a student through behavioral cloning and DAgger. (c) The unified student: unified observation, skill embedding, masked transformer, actor MLP. (d) Transition adaptation: a LoRA on the frozen student trained with target reward from previous-skill and native start states.
    Method overview. Open full size
    01

    Multiple skill experts

    Each skill is learned by its own expert, with a motion prior from reference motions, task rewards and a skill-specific curriculum. Shooting is trained for accurate command tracking, dribbling for stable ball control.

    02

    Multi-skill distillation

    A unified observation layout maps every expert's body, ball and command inputs into one space. Behavioral cloning initializes a skill-conditioned transformer student; dataset aggregation (DAgger) refines it with expert labels on the states the student visits.

    03

    Unified transition adaptation

    With the student frozen, an online, shared transition adaptation method using a context-conditioned low-rank adapter, enabling joint adaptation across multiple skill pairs without retraining individual experts.

    Another body

    Dribbling on a Unitree G1.

    A simulated G1 humanoid dribbling, rendered in the same studio.

    Unitree G1 · Dribbling
    Unitree G1 · Dribbling

    Citation

    @inproceedings{messi,
      title     = {MESSI: Multi-Expert Soccer Skill Integration
                   for Physics-Based Character Control},
      author    = {TODO},
      booktitle = {TODO},
      year      = {TODO}
    }