Trajectory World Models for Heterogeneous Environments (ICML 2025)

This is the official code base for the paper Trajectory World Models for Heterogeneous Environments.

Give it a star 🌟 if you find our work useful!

🔥 News

• 🚩 2025.08.18: The UniTraj Dataset is now available on Huggingface.
• 🚩 2025.06.05: Training code is released.
• 🚩 2025.05.01: TrajWorld has been accepted by ICML 2025, congrats!.
• 🚩 2025.02.03: Our paper is released on arXiv.

🛠️ Installation

conda env create -f environment.yaml
conda activate trajworld

🌟 Pre-training

To pre-train the TrajWorld model:

python scripts/pretrain/pretrain_trajworld.py history_length=20 log_root_dir=log_pretrain_trajworld exp_name=merge_all n_blocks=6

See more baseline scripts in scripts/pretrain.

🎇 Fine-tuning

To fine-tune TrajWorld:

python scripts/training/train_trajworld.py env_name=hopper-medium-replay-v2 log_root_dir=log_model_new trm_epoch_steps=5000 dynamics_max_epochs_since_update=300 dynamics_max_epochs=50 seed=183 train_model_only=true exp_name=trajworld_ft trm_lr=1e-5 load_pt_dynamics_path="pt_model/trajworld_pt/trajworld_pt.pkl" n_blocks=6

To train from scratch, simply remove the load_pt_dynamics_path argument.

More fine-tuning scripts for baseline models can be found in scripts/training.

📊 Evaluation

🔁 Transition Prediction Evaluation

python pred/pred_mse_trajworld.py --env walker2d-random-v2 --model_path <path_to_your_model> --n_blocks 6

Additional baseline scripts: pred.

📈 OPE Evaluation

python ope/ope_eval.py --algo trajworld --env halfcheetah-expert-v2 --clear_kv_cache_every 10 --trm_lookback_window 10 --group 0 --n_blocks 6

Specify your model path by modifying get_list_dirs() in ope/ope_eval.py.

Example commands for other baselines are provided at the top of the file.

🤖 MPC Evaluation

python mpc/mpc.py --algo trajworld --env walker2d-medium-replay-v2 --group 5 --clear_kv_cache_every 10 --trm_lookback_window 10 --action_proposal_id 3 --std 0.1

As above, modify get_list_dirs() in ope/ope_eval.py to set your model path.

Baseline examples are included in the script.

🚀 Release Progress

• [x] UniTraj Dataset
• [x] Pre-trained TrajWorld Model
• [x] Transition prediction evaluation
• [x] OPE evaluation
• [x] MPC evaluation
• [x] Training code for Trajworld, TDM and MLP-Ensemble

📜 Citation

If you find this project useful, please cite our paper as:

@article{yin2025trajectory,
  title={Trajectory World Models for Heterogeneous Environments},
  author={Yin, Shaofeng and Wu, Jialong and Huang, Siqiao and Su, Xingjian and He, Xu and Hao, Jianye and Long, Mingsheng},
  journal={arXiv preprint arXiv:2502.01366},
  year={2025}
}

🤝 Contact

If you have any questions, please contact yinshaofeng04@gmail.com.

💡 Acknowledgement

We sincerely appreciate the following github repos for their valuable codebase we build upon:

• https://github.com/nissymori/JAX-CORL