Teaching a model to watch breaking
Read the full storyA breaking battle moves faster than anyone can type.
Producers spent hours tagging what happens in every clip. I led Red Bull’s first video annotation pipeline: I built a movement ontology with professional breakers, led the research team that fine-tuned a video language model, and kept a person in the loop for the moves the model wasn’t sure of. In a pilot on the Red Bull BC One livestream, it was 87% accurate across 5,100+ labelled clips. Then we hit a research ceiling, and we chose not to scale it.
For product I took it from a proposal to a pilot: the ontology, the data, the evaluation, the partners and the funding.
For innovation Red Bull’s first annotation pipeline, designed so other sports could follow.
For teaching A person correcting a model is the clearest way I know to show how AI learns. Try it below.
- Applied research, compute, and engineering from Google, Microsoft, IBM, Twelve Labs, and Eluvio AI, with NVIDIA hardware processing camera data on site
- Our paper, Breakdance Video Classification in the Age of Generative AI, is under review
Try it: be the person in the loop
A minute of a breaking battle, cut into moves. Press play and watch the model label each move as it happens. Then decide the ones it isn’t sure of.
- 0:00Toprock
- 0:07Toprock
- 0:13Footwork
- 0:21
Freeze?, Check
- 0:26Power move
- 0:34
Footwork?, Check
- 0:43Power move
- 0:51
Freeze?, Check
Illustration. An invented round drawn as a skeleton: not Red Bull footage or model output.












