[Multi robot] save and restore individual robot states for batched RL simulation

[Danfoa]

Hi,

I am currently working with a Legged Locomotion environment in which I have multiple robots running within the same physics client. The multiple robot instances are used in order to use batched inference of the policy network, exploiting the GPU appropriately.

However, I am having a lot of trouble finding out how to restore individual robot states, especially when the state was saved during contact with the ground. When I try to restore each robot state through resetBasePositionAndOrientation and resetJointStates, the robots that were in contact get shot off to the sky (contact force is huge).

Should I keep a bullet client for each robot separate? or should I use a single bullet client with multiple robots? which is possible/more efficient?