search

nv-tlabs / ASE

Other
793 stars 128 forks source link

How to reduce observations to make them more realistic? #45

Closed Robokan closed 1 year ago

Robokan commented 1 year ago

The AMP agent (NN) receives the following observations: root_h_obs, root_rot_obs, local_root_vel, local_root_ang_vel, dof_obs, dof_vel, flat_local_key_pos

On a real physical robot: root_h_obs, local_root_vel, local_root_ang_vel, flat_local_key_pos would probably not be possible to get.

More likely: root rotation, DOF positions, DOF velocities, foot pressure would be the only ones available. From what I can tell AMP compares these observations to what it sees from the motion capture. Is there a easy way to send a reduced set of observations to the agent that matches more of what is observed in reality? Will the algorithm still work?

Robokan commented 1 year ago

After some deeper digging it appears that the AMP observations are only used by the discriminator. So I just changed the observations in humanoid.py and it works great.