Robots that use simulations to plan actions often fail because camera views and sensor readings are imperfect, causing mismatches between the real world and the simulated model. These mismatches lead to errors in precise tasks like pressing small buttons or moving sliders.
To solve this, researchers built ActiveAlign, which lets the robot actively move its hand to calibrate its camera (waving) and then touch objects to adjust their positions with tactile feedback (touching). This improves the match between real and simulated environments.
On a real robot, ActiveAlign successfully performed fine manipulation tasks such as pressing buttons and moving sliders. By actively perceiving through movement and touch, the robot achieved precise alignment, making simulation-to-reality transfer more reliable for delicate operations.