Guidelines to work with mjlab ?

[apirrone]

Hello !

This is very cool, I managed to put my robot in mjlab and make it walk in sim using the velocity env very quickly :)

aze.mp4

But I'm a little confused on what is the "right" way of going deeper and customize things, add/remove observations, rewards, domain randomization, noise etc.

For example, I wanted to train without the  base_lin_vel observation.

Here, I could just comment the term, but I'd like to "extend" from the base tools you give us. I can't really inherit from ObservationCfg because I'd like to remove a term, so I have to duplicate all of ObservationCfg in my ObservationNoLinVelCfg, but now I can't use LocomotionVelocityEnvCfg because observations expects the type ObservationCfg, not ObservationNoLinVelCfg, so I have to duplicate LocomotionVelocityEnvCfg too.

Am I missing something here ? I'm not very familiar with dataclasses yet I probably am !

Could you explain a little the workflow you imagined when designing mjlab ?

Thanks !

[louislelay]

Hi @apirrone, thank you for posting that. Super cool to see your Open Duck Mini walking in mjlab!

How to remove a term?

If you only need to drop one or two terms, do this after super():

@dataclass
class OpenDuckRoughEnvCfg(LocomotionVelocityEnvCfg):
  def __post_init__(self):
    super().__post_init__()
    # your code ...
    self.observations.policy.base_lin_vel = None
    self.observations.critic.base_lin_vel = None

How to add terms?

Inherit from ObservationCfg (e.g., ObservationCustomCfg), then override the env’s observations field:

@dataclass
class OpenDuckRoughEnvCfg(LocomotionVelocityEnvCfg):
  observations: ObservationCustomCfg = field(default_factory=ObservationCustomCfg)
  def __post_init__(self):
    super().__post_init__()
    # your code ...

What if you've much more changes to do?

That’s when you’d create your custom observations cfg from scratch (not inhereting) and, as above, override the env’s observations field to use it.

Please note that this also applies to other managers (rewards, curriculums, etc).

Warning

An upcoming change in #124 will change what I just wrote.

[apirrone]

Thanks a lot @louislelay !