This tutorial: ompl interface tutorial hints that cartesian planning is much faster because IK can be used as a generative sampler, whereas if joint space planning is enforced, rejection sampling is used. I was hoping someone could elucidate where this speed up comes from.
It seems to me that collision detection (the bottleneck for speed) has to be done the same way. Furthermore, looking into the code itself (pose_model_state_space.cpp), nothing looks too fundamentally different in the pose based state space. Sampling still actually happens in joint space.
Any help provided would be much appreciated!