์ ์: Nikhil Sobanbabu, Guanqi He, Tairan He, Yuxiang Yang, Guanya Shi | ๋ ์ง: 2025-05-20 | URL: https://arxiv.org/abs/2505.14266 📄 PDF
Figure 2: Overview of SPI-Active. Data Collection: Collect real-world trajectories using RL policies or
SPI-Active๋ legged robot์ ๋ฌผ๋ฆฌ ํ๋ผ๋ฏธํฐ๋ฅผ ์ํ๋ง ๊ธฐ๋ฐ์ผ๋ก ์๋ณํ๊ณ Fisher Information ์ต๋ํ๋ฅผ ํตํ active exploration์ผ๋ก sim-to-real ๊ฐญ์ ์ต์ํํ๋ two-stage ํ๋ ์์ํฌ์ด๋ค.
Figure 1: SPI-Active enables high-fidelity Sim-to-Real transfer across diverse locomotion tasks. To highlight
Figure 2: Overview of SPI-Active. Data Collection: Collect real-world trajectories using RL policies or
์ดํ: ์ด ๋ ผ๋ฌธ์ legged robot์ sim-to-real ๊ฐญ ํด๊ฒฐ์ ์ํ ์๋ฆฌ์ ์ด๊ณ ์ค์ฉ์ ์ธ system identification ํ๋ ์์ํฌ๋ฅผ ์ ์ํ๋ฉฐ, Fisher Information ๊ธฐ๋ฐ active exploration ์ ๋ต์ ์ฐฝ์์ ์ ์ฉ์ผ๋ก ๊ณ ์ ๋ฐ locomotion ์์ ์์ ํ์ ํ ์ฑ๋ฅ ํฅ์์ ๋ฌ์ฑํ๋ค.