Real-robot results verified: 85% (17/20) vs 25% control, p=0.00033

same run, 10k-step checkpoint

ACT sim-pretrained + 33% real co-train, 20k steps

initial commit