A post calls it one of FAR's most surprising recent projects and links to the collaborators' work.
A post highlighting a recent FAR project says naive discrete actions can be more precise than continuous actions for robot control. The author calls the project surprising and points to work by collaborators.
7.4K
8 Sources, first seen ago
@ZejiYiCommon belief: Robot policies need continuous action heads; discrete actions are π₯ππ¬π¬ π§πππ’π―π ππ§π π₯ππ¬π¬ π©π«πππ’π¬π. βWe find policies w/ the simplest discrete representations π¦ππππ‘ π¨π« ππππ continuous onesβin sim & real, in single-step & diffusion/flow settings. Here's why π§΅
@peterxichenThis is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continuous actions for robot control. Check out this cool work by @ZejiYi et al
@GuanyaShiRT @peterxichen: This is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continβ¦@ZejiYiCommon belief: Robot policies need continuous action heads; discrete actions are π₯ππ¬π¬ π§πππ’π―π ππ§π π₯ππ¬π¬ π©π«πππ’π¬π. βWe find policies w/ the simplest discrete representations π¦ππππ‘ π¨π« ππππ continuous onesβin sim & real, in single-step & diffusion/flow settings. Here's why π§΅@peterxichenThis is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continuous actions for robot control. Check out this cool work by @ZejiYi et al@GuanyaShiRT @peterxichen: This is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continβ¦@ZejiYiCommon belief: Robot policies need continuous action heads; discrete actions are π₯ππ¬π¬ π§πππ’π―π ππ§π π₯ππ¬π¬ π©π«πππ’π¬π. βWe find policies w/ the simplest discrete representations π¦ππππ‘ π¨π« ππππ continuous onesβin sim & real, in single-step & diffusion/flow settings. Here's why π§΅@peterxichenThis is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continuous actions for robot control. Check out this cool work by @ZejiYi et al@GuanyaShiRT @peterxichen: This is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continβ¦@ZejiYiCommon belief: Robot policies need continuous action heads; discrete actions are π₯ππ¬π¬ π§πππ’π―π ππ§π π₯ππ¬π¬ π©π«πππ’π¬π. βWe find policies w/ the simplest discrete representations π¦ππππ‘ π¨π« ππππ continuous onesβin sim & real, in single-step & diffusion/flow settings. Here's why π§΅@peterxichenThis is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continuous actions for robot control. Check out this cool work by @ZejiYi et al@GuanyaShiRT @peterxichen: This is one of the most surprising projects FAR took on recently: (naive) discrete actions can be more precise than continβ¦β
Not ranked yet
β
Not ranked yet
