Learning geometry from vision for 

robotic manipulation