์ ์: Chong Zhang, Victor Klemm, Fan Yang, Marco Hutter | ๋ ์ง: 2026-01-13 | URL: https://arxiv.org/abs/2601.08485 📄 PDF
Fig. 1. Our method enables agile and generalized legged locomotion across diverse terrains with onboard sensing and comp
AME-2๋ Attention ๊ธฐ๋ฐ ๋งต ์ธ์ฝ๋๋ฅผ ํตํฉํ ํตํฉ RL ํ๋ ์์ํฌ๋ก, ๋ฏผ์ฒฉ์ฑ๊ณผ ์ผ๋ฐํ๋ฅผ ๋์์ ๋ฌ์ฑํ๋ ์ฌ์กฑ/์ด์กฑ ๋ก๋ด ๋ณดํ ์ ์ด ๋ฐฉ๋ฒ์ด๋ค. ํ์ต ๊ธฐ๋ฐ์ ๋ถํ์ค์ฑ ์ธ์ elevation mapping ํ์ดํ๋ผ์ธ๊ณผ teacher-student ํ์ต ์ฒด๊ณ๋ฅผ ํตํด sim-to-real ์ด์ ์ ๊ฐ์ ํ๋ค.
Fig. 1. Our method enables agile and generalized legged locomotion across diverse terrains with onboard sensing and comp
Fig. 2. An overview of our system. We use RL to train a teacher policy with ground-truth mapping in simulation, and a st
์ดํ: AME-2๋ Attention ๊ธฐ๋ฐ ๋งต ์ธ์ฝ๋์ ๋ถํ์ค์ฑ ์ธ์ elevation mapping์ ํตํด agile๊ณผ generalized ๋ณดํ์ ํตํฉ์ ์ผ๋ก ๋ฌ์ฑํ๋ ์ฐ์ํ ํ๋ ์์ํฌ์ด๋ฉฐ, quadruped๊ณผ biped ์์ชฝ์์ ์ค์ฆ๋ ๊ฐ๋ ฅํ ์ผ๋ฐํ ๋ฅ๋ ฅ๊ณผ sim-to-real ์ด์ ํจ๊ณผ๋ฅผ ์ ์ฆํจ์ผ๋ก์จ legged locomotion ๋ถ์ผ์ ์ค์ํ ๊ธฐ์ฌ๋ฅผ ํ๋ค.