Actuators vs Predictions

[SirensAI]

Hey all,

I was trying to use an actuator as a "prediction" or the output of the NN. I have 4 control points that I want to predict, the input layer would be a depth camera, using the CartpoleDepthCameraEnvCfg as a basis. Ground truth being the known control points which would act as the reward function. So greater the distance between current and prediction greater the error, its a continuous action space so planning on using Twin Delayed DDPG/TD3 for this.

I had planned to use the actuators as the prediction using 12 DOF on prismatic joints XYZ for each along each axis. As I would need to use rigid bodies to connect the joints I dont think this is the right way to go as its only a prediction in space and not a physical object.

An example would be 4 prims lerping continuously between random 3D points, each prims current position would be the ground truth the observation is the depth camera array. I want to predict the continuous position of each prim.

How could I make the predictions in isolation rather than an articulation command?

[RandomOakForest]

Thank you for posting this. Typically, a direct workflow will provide more flexibility to implement such an environment.