Datasets:
Imitation learning results
Hi, I am also working with the UR5e to evaluate the current imitation learning strategies. Did you perform imitation learning with this task, and if so, how did it perform? I am asking because I am having difficulty generalizing a simple pick and place task (like the one in this dataset). I am using LeRobot and trained an AcT model to do a pick and place, but it always tries to pick the object in roughly the same location. Have you achieved better results the with this dataset?
Lastly, I also noticed that you used the GELLO setup for tele-op. I am doing the same :) If you are up for it, I would like to get some of your feedback on this setup (GELLO + UR5e) -- it seems like we are on similar journeys (or you were).
Hi Zico.
it always tries to pick the object in roughly the same location
I have encountered similar issues as well. In my experience, it often points to either 1) a lack of diversity in the training data (i.e. did you randomize the position of the block significantly?) or 2) not training long enough (where the policy has learnt this open-loop pick as a behavior that somewhat minimizes the loss). I don't recall the exact results from this dataset, but I believe it worked fairly well, though I had to collect way more demonstrations than I expected initially.
I would like to get some of your feedback on this setup (GELLO + UR5e)
It works quite well for me. I did have some issues initially with timings, but other than that we are quite happy with them. The codebase I use to teleoperate with them (which was partially taken from the original GELLO codebase) is available here: https://github.com/tlpss/robot-imitation-glue
As a final note, I always find it useful to initially learn a very simple task with 2D action space, such as pushing the cube to a target location. This helps to check if your codebase does not contain bugs, your hyperparameters are somewhat decent, and your expectations of what to expect from the policy for a certain amount of demonstrations are properly calibrated ( don't expect tooo much from it ;) )
Good luck with your project!
Hi Thomas,,
Thanks for the feedback. I really appreciate it :)
Do you remember the issues you had with timing? Thanks for sharing your repo, I will definitely take a look at this.