clvrai / furniture

IKEA Furniture Assembly Environment for Long-Horizon Complex Manipulation Tasks
https://clvrai.com/furniture
MIT License
502 stars 56 forks source link

Questions Regarding the Repo #35

Open mihdalal opened 2 years ago

mihdalal commented 2 years ago

Hi, I am very interested in using this environment for my research. I have the following questions:

  1. Is this repo still being developed/used? It seems the last commit was from last year so I just want to be sure everything is up to date.
  2. Is there a plan to switch to deepmind's mujoco backend (instead of mujoco py)
  3. Are there any plots with baseline results for RL/IL methods?
  4. Have any of the proposed follow ups from the paper been implemented? (mentioned at the end of paper such as realistic tool attachment or additional robot support).
youngwoon commented 2 years ago

Glad to hear that you're interested in our environment.

  1. We have been working on this environment, but haven't completed it yet. For a pre-mature version, please refer to tstar branch.
  2. We don't have a plan to switch our MuJoCo backend. Is there any reason you want dm_control over mujoco_py?
  3. You can find learning curves of baseline methods in our ICRA paper: https://clvrai.com/assets/research/lee_icra21.pdf
  4. We have been mainly working on improving usability so far.
mihdalal commented 2 years ago
  1. Can you describe what is the difference between master and tstar? If i was to start a new research project, I would ideally want something that isn't changing too much but also has working functionality with well tuned baselines.
  2. Mujoco_py hasn't been updated in years and will likely become deprecated at some point, especially given that DeepMind is now re-writing mujoco. Additionally, mujoco_py has many annoying setup and headless rendering issues that will never be fixed at this point. My suggestion was to use: https://github.com/deepmind/mujoco (not dm_control) since deepmind is actively improving it. I totally understand that this is a ton of work though so this was more to just bring it your attention.
  3. Thank you for the baseline curves! I would suggest adding this to the website, currently the website links to the old paper from 2019 which is heavily out of date it seems. :)
  4. Sounds good!
youngwoon commented 2 years ago

Thank you for your detailed answers.

  1. The tstar branch includes improved dense reward functions and a scripted expert policy for demo collection.
  2. I don't have experience with DeepMind's MuJoCo python binding but it looks well maintained. I'll take a closer look and think about replacing mujoco_py with DeepMind's MuJoCo python binding. Thank you for your suggestion!
  3. Sure. Will update the website together with the future code update.