How to build task embedding in Isaac Lab for long horizon tasks?

[SouvickC]

Is CommandManagerCfg a good place to build the task embedding?

[Miko997]

Depends on what you mean by embedding.

If it's a task ID, goal vector or current subtask, a custom CommandTerm is a reasonable place. The config sets it up; the term holds and updates the per-environment state. You can expose that to the policy through an observation using mdp.generated_commands.

For a long-horizon task, I'd make the phase transitions explicit and reset the phase state for the relevant environment IDs.

If it's a learned embedding of observations or history, I'd keep the encoder in the policy/training model instead. The command manager supplies the task information; the model learns its representation.

Are you thinking of a fixed task vector or a latent learned alongside the policy?