[Proposal] Gymnasium/MuJoCo-v5 dataset generation

[Kallinteris-Andreas]

Proposal

I will make Datasets for the gymnasium/mujoco-v5 environments
The Datasets will be provided in 5 stages for sake of review process

Stage 1

Environments: InvertedPendulum, InvertedDoublePendulum
Datasets: 1 "Medium" + 1 "Expert" (the "expert policy" here mean "perfect", because the environments are so simple)

Stage 2

Environments: Reacher, Pusher
Datasets: 1 "Medium" + 1 "Expert"

Stage 3

Environments: Hopper, Walker, HalfCheetah
Datasets: 1 "Simple" + 1 "Medium" + 1 "Expert"

Stage 4

Environments: Swimmer (this is a bit of special case cause of the learning rate requirement)
Datasets: 1 "Expert"

Stage 5

Environments: Ant, Humanoid, Humanoid Standup
Datasets: 2 "Simple" + 2 "Medium" + 1 "Expert" (note: the expert here will be a fine-tuning policy, like we have Ant-v5 already)

Stage 6 (Bonus maybe?)

Environments (from menagerie real robots): "GoogleBarkour-v1", "Unitree-G1"
Datasets: (probably) 1 "Medium" + 1 "Expert"

Motivation

This is already well requested.

Pitch

Alternatives

Any suggestion welcome, or alterations to current plan are welcome

Additional context

The Dataset generation process will begin after gym1.0 is released.
I am the project manager of Gymnasium Robotics

The Rodrigo's code (https://github.com/Farama-Foundation/minari-dataset-generation-scripts/pull/8/files) will likely be adapted for the new versions of SB3, GYM, MINARI, rather than writing it from scratch

Checklist

• [] I have checked that there is no similar issue in the repo (required)

[younik]

Perfect, this sounds good, thanks!

I suggest to use MuJoCo as namespace for the datasets

[jamartinh]

Hi, please add the excluded variables of observations in infos please as in D4RL.

[Kallinteris-Andreas]

@jamartinh waiting on #250

[younik]

Why there are excluded variables in the observation? @Kallinteris-Andreas

[Kallinteris-Andreas]

Example:
x- and y-coordinates of the Ant's torso (https://gymnasium.farama.org/environments/mujoco/ant/#observation-space)

He is referring to the fact that info is not being included in datasets

[jamartinh]

@younik

With respect to #250

You have to decide, accept the proposal or reject it. I don't see any other option. I have proposed a solution and a default behavior. Or either reject the enhancement for the user in favor of design and expectation simplicity or adopt only the new format, so:

1. Keep the enhancement as optional and defaulting to current behavior Support heterogeneous dicts in infos #250
2. Reject the enhancement Support heterogeneous dicts in infos #250
3. Adopt the enhancement Support heterogeneous dicts in infos #250 as the new way for now on

This is up to you.

[younik]

Example: x- and y-coordinates of the Ant's torso (https://gymnasium.farama.org/environments/mujoco/ant/#observation-space)

He is referring to the fact that info is not being included in datasets

Oh, got it, thanks!

@younik

With respect to #250

You have to decide, accept the proposal or reject it. I don't see any other option. I have proposed a solution and a default behavior. Or either reject the enhancement for the user in favor of design and expectation simplicity or adopt only the new format, so:

1. Keep the enhancement as optional and defaulting to current behavior [Support heterogeneous dicts in infos #250](https://github.com/Farama-Foundation/Minari/pull/250)

2. Reject the enhancement   [Support heterogeneous dicts in infos #250](https://github.com/Farama-Foundation/Minari/pull/250)

3. Adopt the enhancement [Support heterogeneous dicts in infos #250](https://github.com/Farama-Foundation/Minari/pull/250) as the new way for now on

This is up to you.

The fourth option is to implement it in a backward-compatible way, as I explained in the PR comment.
So for now, I will just reject #250. I don't have an ETA for this feature, but maybe @gabrielemaraglino will have some bandwidth to work on it in the next months.