Event

ActiveAlign: Active Perception Alignment for Real2Sim2Real

1 reports · 1 sources

Ri-Zhao Qiu and co-authors propose the ActiveAlign framework, which uses active perception to narrow the Real2Sim2Real gap in egocentric manipulation. The framework contains two modules: ActiveAlign-Waving uses active hand movements and differentiable rendering for hand-eye calibration; ActiveAlign-Touching refines object alignment through tactile feedback, achieving millimeter-level precision. The authors demonstrate on a real Unitree G1 robot, stating that ActiveAlign can perform fine manipulation of millimeter-scale objects such as pressing buttons and moving sliders.