Demystifying Diffusion Policies: Action Memorization and Simple Lookup Table Alternatives

This repository is the implementation of Demystifying Diffusion Policies: Action Memorization and Simple Lookup Table Alternatives. The paper is currently under review. This branch contains the basic version, trained on 30 trajectories. A more advanced version using 120 trajectories is available in the "multimodel" branch. Additionally, we provide code for training a generative model from 2D points uniformly distributed on 4 different shaped 1D manifold.

Testing

To use our pretrained model, you can run the following command:

python rollout.py

We have two prepared virtual tests, InD_cases and OoD_cases, which are different types of tests respectively.

Training

If you want to train the model, you can run the following command:

python training.py

after which you can train the model "fusion_encoder" with the default hyperparameters. Then, run the following command to get the embedding of the raw data:

python data_creation.py

You will get the "traj_database.pt" file for further testing. Run the following command to test the model:

python rollout.py

In the real robot experiments, you will also need to use the function "realtime_rollout" in rollout.py to get the action from the model.

Reference

@article{he2025demystifying,
  title={Demystifying Diffusion Policies: Action Memorization and Simple Lookup Table Alternatives},
  author={He, Chengyang and Liu, Xu and Camps, Gadiel Sznaier and Sartoretti, Guillaume and Schwager, Mac},
  journal={arXiv preprint arXiv:2505.05787},
  year={2025}
}