search

leggedrobotics / wild_visual_navigation

Wild Visual Navigation: A system for fast traversability learning via pre-trained models and online self-supervision
https://sites.google.com/leggedrobotics.com/wild-visual-navigation
MIT License
120 stars 9 forks source link

Traversability value saturation #310

Open Matthewjsiv opened 1 month ago

Matthewjsiv commented 1 month ago

Thanks for your work on open sourcing this and making it easy to run! I just had a quick clarification question regarding the traversability supervision signal.

I noticed on both of the outdoor bags that the value published to "/wild_visual_navigation_node/instant_traversability" are generally saturated in the range of .98-1.0. I am wondering if this is expected or I have something set up wrong on my end, as in the paper it is mentioned that the values are calibrated based on the platform to stretch between 0-1.0.

Additionally, I was wondering if you guys are planning on releasing anymore bags with more diverse robot-environment interactions, for example the experiments in Figs 1&11 from the 2024 paper.

Thanks again!

mmattamala commented 1 week ago

Hi @Matthewjsiv, apologies for the very late reply. If I remember correctly, the sequences in the rosbags are not hard enough to be make that distinguishable. If you want to make it more sensitive, you can change the parameters of the sigmoid function, which unfortunately we did not expose in the config files: https://github.com/leggedrobotics/wild_visual_navigation/blob/8b9caf915e668653b761b57ed33d2719a38b4bfd/wild_visual_navigation_ros/scripts/wvn_learning_node.py#L95-L106

Regarding the rosbags, we are not planning to publicly release more bags at the moment. There are other ongoing projects (especially on @JonasFrey96's side) which involve more interesting data to be released but we don't have a timeline for this I'm afraid.