Closed-loop gripper simulation


I’m using a self-created robot arm model to run some reinforcement learning algorithms and I find my gripper is unstable which has closed-loop constraints. After model .reset(), it can go wrong like the figures I show below and sometimes, after some iteration, it manages to self-correct.

So I wonder if there is any suggested way to fix this problem while maintaining the structure like the physics robot?

PS: I’m using the OmniIsaacGymEnvs repo to clone the robots and do all the trainings.


The USD model screenshots with open gripper status and closed one:


Please have a look at this topic: Closed articulation simulation problem - #28 by gyeoman

We are aware of some suspected issues with GPU simulation and closing articulation loops that we are looking into. But please try PGS, and lower iteration counts coupled with smaller time steps as @gyeoman recommends in that thread and see if this helps.