aysin/sparse_medium_expert_D4RL

Sparse Medium-Expert D4RL

Sparse versions of the D4RL Gym-MuJoCo medium-expert datasets (HalfCheetah, Hopper, Walker2d), introduced in Generative OOD-regularized Model-based Policy Optimization (GORMPO).

Description

These datasets test offline RL methods under thin state-action coverage, where policies are prone to drifting into out-of-distribution (OOD) regions. For each dataset, we define an "unsafe" region that encloses one mode of the action-norm vs. reward space, and randomly discard 50% of the trajectories passing through it. Overall, this removes about 27.5% (HalfCheetah), 22% (Hopper), and 27% (Walker2d) of the original samples.

We chose the medium-expert level because its mix of behavior quality resembles clinical decision-making data.

Citation

@inproceedings{tumay2026gormpo,
  title     = {Generative OOD-regularized Model-based Policy Optimization},
  author    = {Tumay, Aysin and Huang, Jiahe and Jortberg, Elise and Yu, Rose},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}