I am in RL but new to robotics. I am trying to know what the best to train an RL policy in an end2end fashion for grasping or manipulations tasks using images. I can think about three ways. Can someone please help me to understand which one of them work in practice and what are the shortcomings of these techniques:

  1. Starting from scratch with CNN + RL: Not practical since training CNN might be tough.
  2. Use pre-trained network (such as resnet) as a feature extractor. Freeze the weights of CNN and train only the policy network.
  3. Use pre-trained network but Don't freeze the weights of CNN. Fine-tune the entire network while learning the policy.

Please let me know if there are any other approaches that I am missing. Which one of them is more popular and provide better results for robotics (grasping and manipulation) tasks. Is there any git repo that I can readily use that includes a simulator and easy to use/ understand codes?

Also, one of the disadvantages of applying RL to end-2-end learning scenario is that we still need a manual crafted reward for each input image. Do you know how is this generally done? I am aware of RL-guided policy search approaches but are you aware on any other approch for reward calculation using images in an end-end scenario?



Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.