[Question] RslRlSymmetryCfg example

[jloganolson]

Has anyone tried using isaaclab_rl.rsl_rl.RslRlSymmetryCfg ?

I couldn't find any examples or documentation beyond this: https://isaac-sim.github.io/IsaacLab/main/source/api/lab_rl/isaaclab_rl.html#isaaclab_rl.rsl_rl.RslRlSymmetryCfg

[bikcrum]

@jloganolson

Yes, I’ve used it. It is fairly straightforward once you understand how the mirroring should apply to your specific robot or actuator configuration.

The key is to implement a data_augmentation_func that returns both the original and mirrored versions of the observations and actions. The mirroring logic itself depends on the symmetry of your robot and needs to be implemented manually.

---

Step 1: Implement the Mirror Functions

def mirror_policy_observation(policy_obs):
    # Your logic to mirror policy observations
    return mirrored_policy_obs

def mirror_critic_observation(critic_obs):
    # Your logic to mirror critic observations
    return mirrored_critic_obs

def mirror_actions(actions):
    # Your logic to mirror actions
    return mirrored_actions

---

Step 2: Define the Data Augmentation Function

def data_augmentation_func(env, obs, actions, obs_type):
    if obs_type == 'policy':
        obs_aug = torch.cat((obs, mirror_policy_observation(obs)), dim=0)
    elif obs_type == 'critic' and obs is not None:
        obs_aug = torch.cat((obs, mirror_critic_observation(obs)), dim=0)
    else:
        raise ValueError(f"Mirror logic for observation type '{obs_type}' not implemented")

    actions_aug = torch.cat((actions, mirror_actions(actions)), dim=0)
    return obs_aug, actions_aug

---

Step 3: Add to PPO Runner Config

@configclass
class MyPPORunnerCfg(RslRlOnPolicyRunnerCfg):
    ...
    algorithm = RslRlPpoAlgorithmCfg(
        ...
        symmetry_cfg = RslRlSymmetryCfg(
            use_data_augmentation=True,
            use_mirror_loss=True,
            mirror_loss_coeff=1.0,
            data_augmentation_func=data_augmentation_func,
        )
    )

---

Let me know if you have any specific questions about this.

[jloganolson]

This was really helpful, thank you! I stood it up with my G1 23dof in a basic locomotion policy (source code here)

One thing I need to dig into further - based on initial results, symmetry loss signficantly improves the policy, but use_data_augmentation introduces new asymmetries/artifacts.

[jloganolson]

Fixed here https://github.com/jloganolson/g1_23dof_locomotion_isaac/blob/ee072967a7bea4cd6c7c89d64b76a421957245da/source/g1_23dof_locomotion_isaac/g1_23dof_locomotion_isaac/tasks/manager_based/g1_23dof_locomotion_isaac/agents/rsl_rl_ppo_cfg.py#L19