ARISE-Initiative
Human Demonstrations link in Doc missing
Hello, When I look at this page, it says the .h5 demos should be in models/assets/demonstrations but in the robosuite repo, there is no folder named like this. Also, when I click datasets collected with robosuite indicated in the document, it shows 404 page, which is strange..! Does anyone know how to download these human demonstrations?
Hello! I believe the demonstrations were removed since they were taking up some space in the repository (we should actually update this in the documentation as well). Instead, I think you can use the Robomimic dataset and convert them to Robosuite compatible format. More information on this can be found here: https://robomimic.github.io/docs/datasets/robomimic_v0.1.html
Thank you for the detailed answer! I was wondering about the part: "convert them to Robosuite compatible format.". Currently, our project needs RGB observations, and I was confused about two different descriptions:
(1) In the document referred from the git issue here (https://robomimic.github.io/docs/datasets/robosuite.html), it says to use dataset_states_to_obs.py (2) In the document about Robosuite here (https://robosuite.ai/docs/algorithms/demonstrations.html), it says to use playback_demonstrations_from_hdf5.
(1) is in Robomimic git repo, and (2) is in Robosuite git repo. I was wondering if it's okay to use just 1, and also wondering, then what is the difference between robomimic and robosuite.
Additionally, the command given in the answer above here (https://robomimic.github.io/docs/datasets/robomimic_v0.1.html) does not seem to work with some Robosuite tasks (ex. 'stack'), so I was wondering how can I get access to the demos of these tasks.
Apologies for the delay: for datasets converted into robomimic format, use robomimic scripts e.g. dataset_states_to_obs.py. For datasets collected directly in robosuite (and not converted to robomimic), use robosuite scripts.
| what is the difference between robomimic and robosuite
Great question: robomimic builds on top of robosuite and has (reinforcement) learning algorithms/policy training/architecture implemented. See this image from robomimic's website
.
Robosuite contains the base environments, teleoperation devices, robot controllers, interface with the underlying mujoco simulator, etc.
See this diagram from the robosuite white paper:
