search

ARISE-Initiative / robomimic

robomimic: A Modular Framework for Robot Learning from Demonstration
MIT License
592 stars 181 forks source link

HBC Algorithm Suggestions #151

Closed Dhanushvarma closed 5 months ago

Dhanushvarma commented 5 months ago

When collecting demonstrations for imitation learning algorithms, is it better to prioritize short/expert trajectories or deliberately collect long/expert trajectories to gather more data points?

Additionally, I'm seeking suggestions from the authors on which hyper parameters to adjust in the HBC algorithm to effectively capture multi-modal behavior. Any insights or guidance would be greatly appreciated. Thank you!

amandlek commented 5 months ago

In my experience, shorter trajectories are usually better.

In terms of HBC, the VAE is very sensitive to the "beta" parameter (the weight for the KL loss) - tuning that carefully is likely the most important parameter.

Dhanushvarma commented 5 months ago

In regards to the horizon, does it have to ve tuned as per the task, or is the recommended horizon of 10 work well across multiple tasks?

amandlek commented 5 months ago

It should work well, but might need tuning depending on the task.