Multi-Scale Control Signal-Aware Transformer for Motion Synthesis without Phase

Synthesizing controllable motion for a character using deep learning has been promising approach due to its potential learn compact model without laborious feature engineering. To produce dynamic from weak control signals such as desired paths, existing methods often require auxiliary information phases alleviating ambiguity, which limits their generalisation capability. As past poses contain useful hints, in this paper, we propose task-agnostic method, namely Multi-scale Control Signal-aware Transformer (MCS-T), with an attention based encoder-decoder architecture discover the implicitly synthesizing explicitly requiring phase. Specifically, encoder is devised adaptively formulate patterns of character's multi-scale skeletons, and decoder driven by further synthesize predict state paying context-specialised encoded patterns. result, it helps alleviate issues low responsiveness slow transition happen conventional not information. Both qualitative quantitative experimental results on biped locomotion dataset, involves diverse types transitions, demonstrate effectiveness our method. In particular, MCS-T able successfully generate motions comparable those generated