The value of crowdsourcing robot demonstrations lies in the project’s ability to handle noisy data. On this front, @axisrobotics’s approach is quite pragmatic. Their pipeline includes: ☆ Quality filtering ☆ Converting noisy human demonstrations into usable trajectories ☆ Removing hesitation or stall motions and applying Savitzky–Golay filtering for trajectory smoothing (not yet explicitly confirmed by Axis on X) ☆ Resampling using cubic splines and temporally aligning robot/object states to a stable 20 Hz control grid Axis also openly acknowledges that raw web demonstrations suffer from micro-jitters, variable frame rates, and other issues—making these raw data insufficient on their own for reliable policy training. They also describe replay validation of submitted trajectories. Not avoiding pain points is a positive sign; what matters next is observing how effective this validation is at scale. Do you think the more important moat is the number of contributors, or the ability to transform raw demonstrations into training-ready trajectories?