์ ์: Zhengyi Luo, Jinkun Cao, Alexander Winkler, Kris Kitani, Weipeng Xu | ๋ ์ง: 2023-05-10 | URL: https://arxiv.org/abs/2305.06456 📄 PDF
Figure 1: We propose a motion imitator that can naturally recover from falls and walk to far-away reference motion, perp
Physics ๊ธฐ๋ฐ humanoid controller์ธ Perpetual Humanoid Controller (PHC)๋ noisy input๊ณผ unexpected falls์ ๊ฐ๊ฑดํ๋ฉด์ 10,000๊ฐ์ motion clips์ ํ์ตํ ์ ์์ผ๋ฉฐ, ์๋ก์ด Progressive Multiplicative Control Policy (PMCP)๋ฅผ ํตํด catastrophic forgetting ์์ด ๋๊ท๋ชจ motion database์์ ํ์ต ๊ฐ๋ฅํ๋ค.
Figure 4: (a) Imitating high-quality MoCap โ spin and kick. (b) Recover from fallen state and go back to reference motio
Figure 2: Our progressive training procedure to train primitives P(1), P(2), ยท ยท ยท , P(K) by gradually learning harder a
์ดํ: ์ด ๋ ผ๋ฌธ์ external force ์ ๊ฑฐ์ PMCP๋ผ๋ novel mechanism์ผ๋ก physics-based motion imitation์ scalability ๋ฌธ์ ๋ฅผ ํจ๊ณผ์ ์ผ๋ก ํด๊ฒฐํ๋ฉฐ, natural fail-state recovery์ noisy input ๊ฐ๊ฑด์ฑ์ผ๋ก ์ค์ video ๊ธฐ๋ฐ avatar application์ ์ฒ์์ผ๋ก ์ค์ฉ์ ์ธ solution์ ์ ๊ณตํ๋ค.