Anywhere at anything by anyone! Their system removes the need for a real robot during demonstration collection, allowing data collection to potentially parallelize without being limited by expensive real robots. Finally, since videos can be captured anywhere, AR2-D2 re-situates demonstration collection out of the laboratory; users can take videos anywhere, making it easy to collect demonstrations involving the manipulation of diverse objects.

串文

2023-09-04 06:28

讚

回覆

轉發

作者

AF
aipaperaf

粉絲

1,122

串文

93+

讚

回覆

轉發

24小時粉絲增長

無資料

互動率

(讚 + 回覆 + 轉發) / 粉絲數

0.09%

回覆 (BETA)

最先回覆的內容
發文後	用戶	內容
幾秒內	AF aipaperaf	How can they get trainable robot data from AR2D2? Once the user records themselves manipulating an object, AR2-D2 extracts the following information: 6D hand pose, hand state, RGB frames and depth estimations. They replace the hand with an AR robot, aligning its motions to align its end effector with the hand. They create a 3D voxelized representation over time from the extracted information. This 3D representation is used to train a PERACT agent.