Closed ebianchi closed 1 year ago
All comments addressed without disagreement.
I additionally opened an issue #7 as a reminder to address the current hacked way of getting experiment-level storage locations, so that the issue is remembered beyond the closing of this PR.
Contents
The main aim of this PR is to add in the ability to choose whether rollout videos render the predicted trajectories with the original geometry or with the current learned geometry. This decision can be toggled via the flag
update_geometry_in_videos
in the configuration classSupervisedLearningExperimentConfig
.Tangential to this main aim, the following other minor changes are also incorporated:
examples/contactnets_simple.py
features settingupdate_geometry_in_videos = True
so reviewers can demo this PR feature.Call for comments
In this implementation, the
MultibodyLearnableSystem
extension of theSystem
class now has an additional (optional) argument for itssummary
method. This argument is the flag for whether to use original or learned geometry in the generated rollout videos. As originally created, this PR does not include any edits to the originalSystem
class or any other extensions of it to maintain a consistent set of arguments forsummary()
. This can be added in if desired, and I welcome feedback on this point in particular.Examples
The following two gifs are of the same ground truth trajectory (in red) with different predicted trajectories (in blue) as the experiment updates the believed object properties. The geometric updates can be observed in these toss videos.