FPMCO decomposes multi-constraint RL into KL-projection sub-problems, achieving higher reward with lower computing than second-order rivals on the new SCIG robotics benchmark.
Results that may be inaccessible to you are currently showing.
Hide inaccessible results
Feedback