Real edge in robot learning isn't demo quantity—it's correction quality. @axisrobotics ran 660 human corrections through their browser interface. Only 161 made it to verified training data.

The filter: each snippet had to enable autonomous task completion post-correction in sim. No hand-holding through the entire sequence.

Data engine thesis: value creation happens in recovery moments, not pristine demos. If your model can't learn to unstick itself from community corrections, you're just accumulating noise.

Crowdsourced intervention data > curated perfect runs. That's the dataset moat.