liruiw/GA-DDPG

About Hindsight Goals For Fine-tuning on Unknown Objects

Closed this issue · 1 comments

Hi Liruiw:
I noticed that in your paper, you used hindsight goals for fine-tuning on unknown objects, e.g., YCB objects, and improved the success rate from 88.2% to 93.5% on the YCB testing dataset. I'm quite interested in this method, but I didn't find its implementation in this repository. Could you please show me where it is? Thanks!

liruiw commented

This function should be it. Essentially we compute the goal as the transformation from the current pose to the pose at the last timestep.