Exploring Different Modes of Transfer for Learning In Hand Manipulation
Exploring Different Modes of Transfer for Learning In Hand Manipulation
Embedded Files
We explored how both transfer to the real world and transfer to a new hand affects trained policy performance on in hand manipulation tasks. We also evaluated how this is different for different state and action spaces.
We explored how both transfer to the real world and transfer to a new hand affects trained policy performance on in hand manipulation tasks. We also evaluated how this is different for different state and action spaces.
Trained Hand
Transferred Hand With Different Fingers
Asymmetric Transferred Hand
We trained RL policies to slide an object to a desired position or orientation using a two fingered gripper. We then applied the policies using one shot transfer to hands with different fingers and to the real world. We found that by directly controlling the angle velocity but not providing the joint angles to the policy as state space we see massive gains in transferability both to other hands and to the real world without losing performance on the trained hand. Our results are presented in the following paper in IEEE RAL.
We trained RL policies to slide an object to a desired position or orientation using a two fingered gripper. We then applied the policies using one shot transfer to hands with different fingers and to the real world. We found that by directly controlling the angle velocity but not providing the joint angles to the policy as state space we see massive gains in transferability both to other hands and to the real world without losing performance on the trained hand. Our results are presented in the following paper in IEEE RAL.
Our work indicates that careful state and action space selection, including hiding part of the state can actually result in a more robust system.
Page updated
Google Sites
Report abuse