I’m guessing inputs are the pose (x y z, rx ry rz) and the outputs are your joints.
Two things: unless you are just trying to solve for the final position, it may be smart to add the initial pose. So if the trajectory matters for purposes of avoiding obstacles then knowing the start position and maybe the obstacles in configuration space could be helpful input to the network. That being said, some sort of sequence model where you encode the goal with an mlp and then a decoder for the trajectory would be cool. Id you just want the final joints then an mlp is fine.
Some people use reinforcement learning in this setting too but I tend to think that’s overkill.
As others have said too: you could always use a traditional algorithm like some variant of RRTs to solve this too as it’s mostly a “done” problem for traditional robotics.
Take a look at flux.jl. Last time I checked, they were building in support. You may have some success if you join that community over pytorch or tensorflow.