BIND links candidate robot actions to 2D image features

A post introducing BIND argues that robot image encoders capture geometry and meaning, while policies built on them can need hundreds of demonstrations for simple pick-and-place tasks and break after small camera shifts. BIND links each candidate action to the image feature where the robot’s end effector would move. The post claims this makes policies more data-efficient and robust to unfamiliar viewpoints and object positions, using RGB input without 3D sensors.