Missing min and max scores in mujoco dataset for normalized score

[28dvd28]

Missing min and max scores in mujoco dataset for normalized score

Hi,

I’m encountering some difficulties when using minari.get_normalized_score() with certain MuJoCo datasets.

For example, when working with mujoco/hopper/medium-v0, I get the following error:

ValueError: Reference score not provided for dataset mujoco/hopper/medium-v0. Can't compute the normalized score.

It seems this happens because the reference scores (e.g., min and max returns) are not defined for this dataset. This is problematic for my use case, as I am developing an RL library and trying to reproduce results from papers that report normalized scores on these MuJoCo environments.

I wanted to ask:

• Is there a recommended way to compute or define normalized scores for these environments?
• Are there official or commonly accepted reference min/max (or random/expert) scores that can be used for MuJoCo datasets where they are missing?
• Alternatively, is there a reason why these scores are intentionally omitted for some Minari datasets?

Any guidance on the intended evaluation protocol would be greatly appreciated.

Kind regards,
Davide Cordioli

[razarcon3]

I'm also facing this issue. Which Mujoco environments were you able to get ref min and max scores from?

[28dvd28]

In the end, I resolved the issue by switching libraries. Instead of using Minari, I moved to DSRL, which is quite similar to D4RL (and therefore conceptually close to Minari). DSRL provides offline datasets for environments from Safety Gymnasium and Bullet Safety Gym, along with built-in utilities to compute normalized performance scores.

It is worth noting, however, that DSRL is specifically designed for safety-oriented offline reinforcement learning, which makes its focus different from D4RL.