Hi, thanks for your impressive work!
I have a quick question regarding the action command inputs. In the paper, you mention that the network takes both the end-effector (EE) target position and a base forward velocity command as part of its observation or command space. How does the policy resolve potential conflicts between these two commands?
For instance, what happens if the EE target is fixed in space (e.g., the arm needs to stay at a specific location), but the base is commanded to move forward at 0.35 m/s? Wouldn’t that make it impossible to satisfy both objectives at the same time? I’m curious how your policy learns to prioritize or reconcile such conflicting commands during training or inference.
Thanks again for the great work!
Hi, thanks for your impressive work!
I have a quick question regarding the action command inputs. In the paper, you mention that the network takes both the end-effector (EE) target position and a base forward velocity command as part of its observation or command space. How does the policy resolve potential conflicts between these two commands?
For instance, what happens if the EE target is fixed in space (e.g., the arm needs to stay at a specific location), but the base is commanded to move forward at 0.35 m/s? Wouldn’t that make it impossible to satisfy both objectives at the same time? I’m curious how your policy learns to prioritize or reconcile such conflicting commands during training or inference.
Thanks again for the great work!