Maybe it's this part of the abstract? >Our approach incorporates a beta policy distribution and a multi-critic architecture to model contact-guided motions, ex

Maybe it's this part of the abstract? >Our approach incorporates a beta policy distribution and a multi-critic architecture to model contact-guided motions, exemplified by a challenging quadrupedal robot skateboard task I'm not an expert on this but maybe someone here can explain it a bit about a beta policy distribution and a multi-critic architecture and how come that is good to model contact-guided motions?