search

mees / calvin

CALVIN - A benchmark for Language-Conditioned Policy Learning for Long-Horizon Robot Manipulation Tasks
http://calvin.cs.uni-freiburg.de
MIT License
309 stars 49 forks source link

[Generate More Data] Can I use the simulator to generate more data? #61

Closed aopolin-lv closed 7 months ago

aopolin-lv commented 7 months ago

Hi! I would like to generate some more data of the existing task, such as push_red_block_left. Can I build the relevant data through the repo?

mees commented 7 months ago

You could generate more data from A) generate more human teleop data by using (undocummented) https://github.com/mees/calvin_env/blob/main/calvin_env/vrdatacollector.py B) an expert policy that you have trained on the original data

COST-97 commented 4 months ago

@mees Hello: Could you provide a demo of an expert policy (e.g. a policy obtained using the RoboFlamingo method) to produce more data? under the same Env A, B, C, and D, under the same language annotations? Thank you!

lukashermann commented 4 months ago

According to the RoboFlamingo Github repo, they uploaded pre-trained models on hugging face .