jayLEE0301 / vq_bet_official

Official code for "Behavior Generation with Latent Actions" (ICML 2024 Spotlight)
https://sjlee.cc/vq-bet/
MIT License
113 stars 12 forks source link

Franka Kitchen Dataset #9

Open ziyadsheeba opened 2 weeks ago

ziyadsheeba commented 2 weeks ago

Hi, thanks for the great work!

I just wanted to understand how The Franka Kitchen dataset was generated. The original dataset was introduced in the Relay Policy Learning paper, but if I parse the raw .mjl files I get 605 trajectories instead of the 566 in your dataset. Also, I would like to know if the dataset used respects physics (i.e the action were re-simulated in using the MuJoCo environment) or you simply used the raw trajectories. Unfortunately I didn't find the code for that here or in the BeT paper.

Thanks in advance!