AI Soccer Coach
A local, real-time coach that watches a training drill and says the right word at the right moment.
- Stop.
- Go.
- Switch.
- Wrong foot.
- Smaller touches.
- Head up.
A coach that reacts during the drill, not after it
The idea
Cameras watch a young player working through a drill. The system follows body position, which foot is used, ball touches, changes of direction and pace, and whether the drill is being done as asked. Then it gives short spoken cues immediately.
Where it starts
A private coordination coach has agreed to let us record lessons. Those recordings become the first real dataset, so we learn what cameras can actually see before designing the rest of the system on paper.
What it isn't
It isn't a video-summary app, a cloud platform or a replacement for a human coach. It's a personal R&D project that runs on our own hardware and works without an internet connection.
Six stages, each proven before the next
Each stage builds on real results from the one before it.
01
Record
Two phones on fixed tripods: a wide 4K/60 master view and a side view focused on feet and body. We record whole sessions continuously and sync the cameras with a clap. A small app uploads each file to a home server and checks every byte arrived intact.
02
Analyze
Review the footage offline. Sync the two cameras, mark where each drill starts and ends, and label the moments a coach would react to.
03
Pose & ball
Track the player's body and the ball frame by frame, and cones where useful. We'll test candidate models (such as RTMPose, MediaPipe and YOLO-family detectors) against the recorded footage and choose based on those results.
04
Events
Turn raw tracking into events: approach cone → plant foot → touch ball → change direction → accelerate. Then compare the expected sequence with what actually happened.
05
Real-time coach
Run the pipeline live. When a rule fires, a short cue plays over a speaker straight away.
06
Field hardware
A portable, battery-powered field computer with a desktop GPU, two or three cameras and a speaker. It will need a proper thermal and weather design, and it will work without an internet connection.
Why the live coach runs on rules, not an LLM
event: left foot touched ball
state: drill expects right foot
rule: wrong foot
action: SAY "Wrong foot."
Timing is the product
A cue only helps if it lands before the next touch. Sending frames to a language model and waiting for a reply adds variable delay on every decision. A rule over structured events runs in a predictable, tiny amount of time.
Same situation, same cue
A deterministic state machine always reacts the same way to the same situation. That makes it testable against recorded footage, easy to debug when it's wrong, and predictable for the player.
No network needed
Fields often have poor coverage. Rules running on local hardware keep working with no signal. An LLM in the loop would make every cue depend on a connection or on a much heavier local model.
Where an LLM helps
An LLM can sit above the real-time loop and handle session analysis, drill planning, explaining mistakes, adapting future sessions and talking with the parent or coach. It shapes the plan but doesn't call each touch.
Stage 1: building the data pipeline
Current milestone: record one complete lesson from two fixed cameras, upload both files to the home server, verify them and organize them into a dataset that's easy to analyze.
Work in progress: an Android recording app that works fully offline and uploads later, and a home-server upload service reachable only over a private network. Every file is checked with a SHA-256 hash before the phone copy can be removed. Analysis starts once the first dataset exists.