Does anyone successfully train the locomotion policy with task space action instead of joint space?

[TharitSinsunthorn]

``I'm want to enhance the level of locomotion training by using task space action instead of joint space for quadruped robot, in this case, I might have a higher controlability from IK guidance from DifferentialIKController

I have tried to apply 4 action terms for each leg just for testing like below, with unitree go2 env.

    # joint_pos = mdp.JointPositionActionCfg(asset_name="robot", joint_names=[".*"], scale=0.5, use_default_offset=True)
    FL_pos = mdp.DifferentialInverseKinematicsActionCfg(
        asset_name="robot",
        joint_names=["FL_hip_joint", "FL_thigh_joint", "FL_calf_joint"],
        body_name="FL_foot",
        controller=DifferentialIKControllerCfg(command_type="position", use_relative_mode=True, ik_method="dls"),
        # body_offset=DifferentialInverseKinematicsActionCfg.OffsetCfg(pos=[0.0, 0.0, 0.107]),
    )

    FR_pos = mdp.DifferentialInverseKinematicsActionCfg(
        asset_name="robot",
        joint_names=["FR_hip_joint", "FR_thigh_joint", "FR_calf_joint"],
        body_name="FR_foot",
        controller=DifferentialIKControllerCfg(command_type="position", use_relative_mode=True, ik_method="dls"),
    )

    RL_pos = mdp.DifferentialInverseKinematicsActionCfg(
        asset_name="robot",
        joint_names=["RL_hip_joint", "RL_thigh_joint", "RL_calf_joint"],
        body_name="RL_foot",
        controller=DifferentialIKControllerCfg(command_type="position", use_relative_mode=True, ik_method="dls"),
    )

    RR_pos = mdp.DifferentialInverseKinematicsActionCfg(
        asset_name="robot",
        joint_names=["RR_hip_joint", "RR_thigh_joint", "RR_calf_joint"],
        body_name="RR_foot",
        controller=DifferentialIKControllerCfg(command_type="position", use_relative_mode=True, ik_method="dls"),
    )

Unfortunately that the result is much worse than using of joint space action. Is there any idea to get improved on this, or it should be the problem of parameters tuning? Also, I'm not sure if my setup is in a proper way or not.

[RandomOakForest]

Thank you for starting this discussion. It would be great to have a good tutorial or how-to about this, but we don't have one yet. Could you elaborate on the error you are measuring? Is there a component of the error that may show a specific pattern of motion that is being under-trained or under-penalized?

[TharitSinsunthorn]

Thank you for your reply! I was able to run the training script successfully, but I'm having issues with the training results - the robot isn't learning to walk or move forward properly. My previous assumption is that, providing the task space would support the training result.

What I've tried so far:

• Started with the same reward settings as the example config (I'm using Unitree Go2) and experimented with tuning the reward weights with diff ik task space
• I also tested my environment config with raw action inputs similar to the approach in tutorials/03_envs/run_quadruped_rl_env.py and the differential IK works fine in isolation

I'm relatively new to RL and am trying to improve the training by switching from joint space to task space control to implement some oscillator model that could help the locomotion learning process. However, I haven't been successful in getting the training to work with this approach yet.

So the point that I'm still trying to figure it out is finding if there is specific considerations or common pitfalls when transitioning from joint space to task space for quadruped locomotion.