-
Notifications
You must be signed in to change notification settings - Fork 2.6k
Open
Labels
enhancementNew feature or requestNew feature or request
Description
Proposal
In the MR #1476, the clipping functionality to individual action terms was introduced. However, there are issues with this implementation:
- The clipping parameter is a dictionary. While this may make sense for joint actions (where keys are the names of the joints), it doesn't mean anything at all for other action terms such as task-space action terms
- The clipping is applied to the processed actions. This isn't how any/many libraries think about for clipping. Clipping is done on the raw actions usually (as done in Clips actions to large limits before applying them to the environment #984). In the case of JointAction, the clipping internally might still make sense but then it should be clipping to the simulation joint limits and must be optional (while clipping of raw actions is a MUST for training stability)
I propose that"
- The clipping of processed actions should be specific to only JointAction (where it may even make sense)
- We should have a separate clipping of raw actions as done in Clips actions to large limits before applying them to the environment #984. We can move that to RSL-RL and other RL wrappers but it is definitely something we need to take care of somewhere.
Checklist
- I have checked that there is no similar issue in the repo (required)
Acceptance Criteria
- The clipping of raw actions happens correctly.
- The clipping inside action terms is properly documented and removed where it isn't applicable.
RandomOakForest, SuomiKP31, rajivswamy and dom-lee
Metadata
Metadata
Assignees
Labels
enhancementNew feature or requestNew feature or request