Question about training estimator

[JSSh1n]

I have carefully reviewed the paper and the code, and I have a question.

Using Move Suitcase as an example, after going through ppo_roa_adapt_est and training the estimator, is the system expected to estimate the object’s position using only the depth image and then reach out to grasp the object accordingly?

When we deployed the model, however, we observed that the robot attempted to grasp the object according to the motion data, regardless of where the object was actually located.

Thank you for your time and consideration.

[EGalahad]

1. The ppo_roa_adapt_est is not tested on real world yet. did you deploy depth map based model? that is very new, would like to see more results from you!
2. "the robot attempted to grasp the object according to the motion data, regardless of where the object was actually located": That's expected since you only have one motion in training. the policy can only generalized to what is seen in training. you can check the randomization range of objects in training, which is typically 10-20cm.