jevPOST
Two-layer Jev decisions enable vision-free robotic hand grasping
Jev first chooses an intent such as grasp, carry or release, then chooses the concrete xyz movement and finger commands, selecting within a predefined finite action space.
Jason Zhu
@GoSailGlobalArchived 2026.09.20Originally posted 2026.09.19
Jason Zhu @GoSailGlobalOriginally posted 2026.09.19
即使Jev没有眼睛👀 也能实现机械手抓取 只要问题定义足够可以 First call → choose an intent 例如: grasp / carry / release 也就是先决定:“现在要做什么?” 第二层: Second call → choose the concrete action 例如: x / y / z movement + finger commands 也就是决定:“具体怎么做?” 而且关键点在于,intents 以及每个 intent 的触发条件,都已经定义在 prompt 里 Jev 并不是让模型自由生成一个动作,而是在一个预先定义好的有限动作空间里选择
This post quotes another post: x.com/i/web/status/2101018760371171420
x.com/gosailglobal/status/2101092040666661234Archived 2026.09.20 11:14