The research direction

Research

Generative rollouts of full tactical scenarios. The diffusion-based simulator models the joint distribution of all on-pitch agents, enabling counterfactual analysis and what-if reasoning at the play level.

That is the vision statement, quoted as written — and it is labeled Research deliberately. No such simulator exists at WithoutBall today. What exists is the machinery that would have to underpin one: a generative model of body movement that can already continue, complete, and re-imagine masked spans of motion.

What runs today: an evaluation workbench

Generative rollouts and rigorous evaluation are the same computation viewed from two directions. When the model completes a masked span of movement, comparing the completion to the ground truth is an evaluation; sampling it forward without ground truth is a simulation. Today that machinery runs exclusively in the first direction — as the workbench behind the evaluation suite, where masked-reconstruction, probing, and clustering results are produced and measured.

Today
Prototype
  • Masked-span completion against ground truth
  • Reconstruction error measured per mask type
  • Probe and clustering harness for the embedding
Research direction
Research
  • Free-running rollouts beyond the observed window
  • Multi-agent tactical scenes
  • Counterfactual what-if queries at play level
One generative backbone, two directions: measured against ground truth today; sampled forward as simulation only when that is earned.

What would have to be true

Three things separate the workbench from a simulator: substantially more data than today’s corpus, validated multi-agent dynamics rather than single-body reconstruction, and evaluation protocols for generated scenarios that do not yet exist anywhere in the field. Until those are earned, simulation stays labeled what it is — research.

The data side of that equation is mapped on the data-scale ladder.