่จปๅ†Šไธฆๅˆ†ไบซ้‚€่ซ‹้€ฃ็ต๏ผŒๅฏ็ฒๅพ—ๅฝฑ็‰‡ๆ’ญๆ”พ่ˆ‡้‚€่ซ‹็Žๅ‹ตใ€‚

zeji yi
@ZejiYi
phd @CarnegieMellon | Intern at @Amazon FAR| Previously @Tsinghua_Uni | Studying Control & Robot Learning
ๅŠ ๅ…ฅ May 2024
380 ๆญฃๅœจ้—œๆณจ    400 ็ฒ‰็ตฒ
Common belief: Robot policies need continuous action heads; discrete actions are ๐ฅ๐ž๐ฌ๐ฌ ๐ง๐š๐ญ๐ข๐ฏ๐ž ๐š๐ง๐ ๐ฅ๐ž๐ฌ๐ฌ ๐ฉ๐ซ๐ž๐œ๐ข๐ฌ๐ž. โ—We find policies w/ the simplest discrete representations ๐ฆ๐š๐ญ๐œ๐ก ๐จ๐ซ ๐›๐ž๐š๐ญ continuous onesโ€”in sim & real, in single-step & diffusion/flow settings. Here's why ๐Ÿงต
้กฏ็คบๆ›ดๅคš