We need to move away from "one policy per task" and towards "one policy for all tasks". One first step would be fixing the current problem where, when the walking policy is given 0s as its cmd, it still moves a lot, just without advancing (stomps without moving in any direction). One idea here would be to have the reward that encourages low dof lin vel be scaled up when the command is close enough to 0.
We need to move away from "one policy per task" and towards "one policy for all tasks". One first step would be fixing the current problem where, when the walking policy is given 0s as its cmd, it still moves a lot, just without advancing (stomps without moving in any direction). One idea here would be to have the reward that encourages low dof lin vel be scaled up when the command is close enough to 0.