@omarsar0 20,000 hours of real manipulation data and no saturation. The scaling curve for real operator data is the one that matters and it's still climbing.
@lukas_m_ziegler The pattern keeps repeating: structured data beats expensive hardware. Force awareness from motor current alone. The data layer matters more than the sensor stack.
@IntEngineering Hardware production is scaling. The harder problem is what happens after deployment: who captures the operational data, and in what structure. That determines which robots actually improve.
Pouring popcorn. Folding a shirt. Stacking blocks. The manipulation demos at #ICRA2026 were genuinely better this year.
But the hard part was never the demo. It's doing the same task 500 times, unsupervised, in a room nobody arranged in advance.
Two runs of the same task. One trains a capable model. The other quietly degrades it. It's full of errors the model would just learn to copy.
Separating the two is what our Standards for Data are for.
We can draw these lines because we run the whole loop: collect the data, train and evaluate models on it, feed what we learn back in.
Better data takes a human willing to hold it to a standard. That's the difference between training a robot and filling a hard drive.