ADM-v2: Any-step Dynamics Model v2

This is the code for the paper ADM-v2: Pursuing Full-Horizon Roll-out in Dynamics Models for Offline Policy Learning and Evaluation in ICLR 2026.

Requirements

To install all the required dependencies:

1. Install MuJoCo engine, which can be downloaded from here.
2. Install Python packages listed in requirements.txt using pip install -r requirements.txt. You should specify the version of mujoco-py in requirements.txt depending on the version of MuJoCo engine you have installed.
3. Manually download and install d4rl package from here.
4. Manually download and install neorl package from here.

Run an experiment

python main.py --env [Env] --env-name [Env name] 

The config files act as defaults for a task. They are all located in config. --env refers to the benchmark, D4RL or NeoRL. --env-name refers to the config files in config/. All results will be stored in the result folder.

For example, run ADM-v2 for policy optimization on hopper-medium-v2 dataset of D4RL benchmark:

python main.py --env d4rl --env-name hopper-medium-v2

Citation

If you find this repository useful for your research, please cite:

@inproceedings{
    adm2,
    author       = {Haoxin Lin and
                    Siyuan Xiao and
                    Yi-Chen Li and
                    Zhilong Zhang and
                    Yihao Sun and
                    Chengxing Jia and
                    Yang Yu},
    title        = {ADM-v2: Pursuing Full-Horizon Roll-out in Dynamics Models for Offline Policy Learning and Evaluation},
    booktitle    = {The 14th International Conference on Learning Representations (ICLR'26)},
    year         = {2026},
    address      = {Rio de Janeiro, Brazil}
}