Online Control of Simulated Humanoids Using Particle Belief Propagation

摘要

We present a novel, general-purpose Model-Predictive Controlrn(MPC) algorithm that we call Control Particle Belief Propagationrn(C-PBP). C-PBP combines multimodal, gradient-free sampling andrna Markov Random Field factorization to effectively perform simultaneousrnpath finding and smoothing in high-dimensional spaces.rnWe demonstrate the method in online synthesis of interactive andrnphysically valid humanoid movements, including balancing, recoveryrnfrom both small and extreme disturbances, reaching, balancingrnon a ball, juggling a ball, and fully steerable locomotion in an environmentrnwith obstacles. Such a large repertoire of movementsrnhas not been demonstrated before at interactive frame rates, especiallyrnconsidering that all our movement emerges from simple costrnfunctions. Furthermore, we abstain from using any precomputationrnto train a control policy offline, reference data such as motion capturernclips, or state machines that break the movements down intornmore manageable subtasks. Operating under these conditions enablesrnrapid and convenient iteration when designing the cost functions.