Sign In

Simulation Distillation: Pretraining World Models in Simulation for Rapid Real-World Adaptation

์ž‘์„ฑ์ž
  • Haebom
์นดํ…Œ๊ณ ๋ฆฌ
Empty

์ €์ž

Jacob Levy, Tyler Westenbroek, Kevin Huang, Fernando Palafox, Patrick Yin, Shayegan Omidshafiei, Dong-Ki Kim, Abhishek Gupta, David Fridovich-Keil

๐Ÿ’ก ๊ฐœ์š”

๋ณธ ๋…ผ๋ฌธ์€ ๋กœ๋ด‡ ํ•™์Šต์—์„œ ์ œํ•œ์ ์ด๊ณ  ํ’ˆ์งˆ์ด ํ˜ผํ•ฉ๋œ ์‹ค์ œ ๋ฐ์ดํ„ฐ์— ๋Œ€ํ•œ ๋น ๋ฅด๊ณ  ์•ˆ์ •์ ์ธ ์ ์‘ ๋ฐฉ๋ฒ•์„ ์ œ์•ˆํ•ฉ๋‹ˆ๋‹ค. ์ด๋ฅผ ์œ„ํ•ด ๋ฌผ๋ฆฌ ์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ๋ฅผ ํ™œ์šฉํ•˜์—ฌ ๋ฐฉ๋Œ€ํ•œ ์–‘์˜ ํ–‰๋™ ์กฐ๊ฑด๋ถ€ ๋กœ๋ด‡ ๊ฒฝํ—˜ ๋ฐ์ดํ„ฐ๋ฅผ ์ƒ์„ฑํ•˜๊ณ , ์ด๋ฅผ ํ†ตํ•ด ํ›ˆ๋ จ๋œ "์›”๋“œ ๋ชจ๋ธ"์„ ์‹ค์ œ ์„ธ๊ณ„์— ์ ์šฉํ•˜๋Š” Simulation Distillation (SimDist) ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์†Œ๊ฐœํ•ฉ๋‹ˆ๋‹ค. SimDist๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜์—์„œ ํ•™์Šต๋œ ์ธ์ฝ”๋”, ๋ณด์ƒ ๋ชจ๋ธ, ๊ฐ€์น˜ ํ•จ์ˆ˜๋ฅผ ์ „์ดํ•˜๊ณ , ์‹ค์ œ ๋ฐ์ดํ„ฐ๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์ž ์žฌ ์—ญํ•™ ๋ชจ๋ธ๋งŒ ์—…๋ฐ์ดํŠธํ•จ์œผ๋กœ์จ ํšจ์œจ์ ์ธ ์‹ค์„ธ๊ณ„ ์ ์‘์„ ๋‹ฌ์„ฑํ•ฉ๋‹ˆ๋‹ค.

๐Ÿ”‘ ์‹œ์‚ฌ์  ๋ฐ ํ•œ๊ณ„

โ€ข
๋กœ๋ด‡ ํ•™์Šต์—์„œ ์‹ค์ œ ๋ฐ์ดํ„ฐ ๋ถ€์กฑ ๋ฌธ์ œ๋ฅผ ๋ฌผ๋ฆฌ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ธฐ๋ฐ˜ ์‚ฌ์ „ ํ›ˆ๋ จ์œผ๋กœ ํ•ด๊ฒฐํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.
โ€ข
SimDist๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜์—์„œ ์–ป์€ ๊ตฌ์กฐ์  ์‚ฌ์ „ ์ง€์‹๊ณผ ์‹ค์ œ ๋ฐ์ดํ„ฐ์˜ ๋น ๋ฅธ ์ ์‘์„ ๊ฒฐํ•ฉํ•˜์—ฌ ํšจ์œจ์ ์ธ ์˜จ๋ผ์ธ ๊ณ„ํš ๋ฐ ๊ฐœ์„ ์„ ๊ฐ€๋Šฅํ•˜๊ฒŒ ํ•ฉ๋‹ˆ๋‹ค.
โ€ข
ํ–ฅํ›„ ์—ฐ๊ตฌ์—์„œ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๊ณผ ์‹ค์ œ ์„ธ๊ณ„ ๊ฐ„์˜ ๋” ์ •๊ตํ•œ ๋ถˆ์ผ์น˜ ๋ฌธ์ œ๋ฅผ ํ•ด๊ฒฐํ•˜๊ณ , ๋‹ค์–‘ํ•œ ๋กœ๋ด‡ ํƒœ์Šคํฌ ๋ฐ ํ™˜๊ฒฝ์— ๋Œ€ํ•œ SimDist์˜ ์ผ๋ฐ˜ํ™” ์„ฑ๋Šฅ์„ ํ–ฅ์ƒ์‹œํ‚ค๋Š” ๊ฒƒ์ด ๊ณผ์ œ๋กœ ๋‚จ์•„์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿ‘