Docs — WithoutBall

Research program

Four pillars, from skeleton to serving.

The stack as designed. Each card opens the full specification — with its current engineering status stated plainly.

Architecture

Hierarchical Motion Transformer

Multi-scale transformer with joint-level graph attention and causal temporal layers. Spatial self-attention resolves inter-joint dependencies across the full skeletal topology, while hierarchical temporal modeling captures movement at multiple granularities — from sub-second micro-actions to full-possession tactical arcs.

Graph AttentionCausal AttentionMulti-Scale

Prototype

Foundation

Football Foundation Model

Pre-trained on movement data at scale — a massive and growing corpus of movement video, converted into skeleton data and spanning playing levels and styles. A single large-scale model that transfers to any downstream task — from injury risk to tactical classification — with minimal fine-tuning.

Foundation ModelTransfer LearningScale

Research

Simulation

Neural Match Simulation

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

World ModelCounterfactualGeneration

Roadmap

Infrastructure

Edge-Optimized Inference

Quantization, compression into compact task-specific models, and hardware-aware graph compilation for deployment at the edge. Asynchronous pipeline with batched pose estimation, streaming transformer inference, and lock-free result aggregation — targeting sub-frame latency from raw video to tactical output.

Model CompressionEdge ComputeAsync Pipeline

Reference

The numbers, the skeleton, the interface.

How the model is measured, the coordinate system it lives in, and the query surface it will expose.

Evaluation

Evaluation Suite

Reconstruction error under occlusion, linear concept probes, and strict held-out validation — every metric published with its qualifiers attached.

Data Model

The 21-Joint Skeleton

Why the platform measures bodies in 21 joints and 20 bones — and how standard pose output maps onto that coordinate system, joint by joint.

Roadmap

Interface

Knowledgebase API

How analysts, tools, and AI agents will query the movement knowledgebase — every answer shipped with pointers back to the underlying movement.