Researchers Comment on Robot In-Context Learning Demo
Andrew Carr suggests 3D simulation for robot training while Yoav Goldberg flags grip and timing flaws in the video.
Andrew Carr responded to Lerrel Pinto's robot video by proposing 3D simulation to remove the need for physical demonstrations. Yoav Goldberg replied that the robot squeezed too hard and released too soon, striking the bowl side. He added that semantic understanding remains limited and that the system likely followed a simple color-based rule of picking yellow and dropping over blue.
Turns out that doing In-context learning for robots is not that hard...
Reactions
2 postsAC 
Andrew Carr ๐คธ@andrew_n_carrnow imagine if you could simulate this in 3D so you didn't need to do the demonstrations yourself!! https://twitter.com/LerrelPinto/status/2093400831941116254
- likes: 36
- replies: 3
- bookmarks: 9
- reposts: 0
(' (((ู()(ู() 'yoav))))๐พ@yoavgoam I the only one seeing the robot squeeze too hard and also dropping too early hitting the side of the bowl? (not saying this is easy, but semantic understanding is not quite there yet --- i assume it learned "pick yellow drop over blue")โฆ
- likes: 4
- replies: 1
- bookmarks: 1
- reposts: 0