Data selection > data volume in robot training. @axisrobotics TUCO study: 10 real demos + 1,200 sim demos per task. Standard lighting, three physical tasks.
Full sim pool (1,200): 25.8% success rate.
Curated 600 demos: 85.8% success rate.
3.3x performance gain from cutting half the data. Real-task feedback drove selection.
Key insight: marginal value of each demo matters more than absolute quality. Combinations that complement existing examples outperform random sampling. Data engines need combinatorial optimization logic, not just volume scaling.
Implication for $AI robotics: training efficiency is a moat. Teams that solve data curation at scale will compress capex and iteration cycles. Worth tracking which labs are building selection infrastructure vs brute-forcing with compute.
Full sim pool (1,200): 25.8% success rate.
Curated 600 demos: 85.8% success rate.
3.3x performance gain from cutting half the data. Real-task feedback drove selection.
Key insight: marginal value of each demo matters more than absolute quality. Combinations that complement existing examples outperform random sampling. Data engines need combinatorial optimization logic, not just volume scaling.
Implication for $AI robotics: training efficiency is a moat. Teams that solve data curation at scale will compress capex and iteration cycles. Worth tracking which labs are building selection infrastructure vs brute-forcing with compute.