Thousands of store cameras record all day for no one. Relentless turns that footage into sorted, labeled, upscaled and multiplied clips for teams training robotics, VLA and world models.
Raw video is the ore. The engine is the refinery. What comes out is data a lab can train on the same day.
A fast first pass keeps only footage with real activity, like hands reaching, picking and placing. Dead hours never reach the expensive stages.
Clips are cut into action segments and labeled ("reaches for item → picks up → places in basket"). Faces are blurred.
Every clip is upscaled to recover resolution and fine motion. We ship the source frames next to the output so you can check it.
Physics-preserving diffusion makes variants of each clip, like new lighting or a swapped object. One real clip becomes many examples.
Stores already have cameras. We plug in our hardware and pay them for footage they were throwing away.
Search by action, setting or object. Every clip comes with action labels, the raw source and the upscaled version.
Pretrain on volume, fine-tune on the premium tier. Pick the mix that fits your model.
Researchers and labs get the first datasets. Store owners can sign up for a box.