Even though it seems like a lot of pointless work, it might speed up RL training and prevent agents from learning unrealistic grasps (currently, the finger contact surface is 3x larger than it should be, so the agent might exploit something that would never work in real life)
Therefore, go through each link and either simplify geometry to improve performance (lower links) or make geometry more representative to improve Sim2Real transfer (hand and finger links)
[x] link0
[x] link1
[x] link2
[x] link3
[x] link4
[x] link5
[x] link6
[x] link7
[x] hand
[x] finger
Results:
It could be optimised much further, but I stopped at this level. In the future, I might add a low-poly version to a separate branch.
Even though it seems like a lot of pointless work, it might speed up RL training and prevent agents from learning unrealistic grasps (currently, the finger contact surface is 3x larger than it should be, so the agent might exploit something that would never work in real life)
Results:
It could be optimised much further, but I stopped at this level. In the future, I might add a low-poly version to a separate branch.