Temporal Difference Learning for Model Predictive Control

(Nov 2023) Announcement: TD-MPC2 is out! Visit https://github.com/nicklashansen/tdmpc2 for more information.

---

Original PyTorch implementation of TD-MPC from

Temporal Difference Learning for Model Predictive Control by

Nicklas Hansen, Xiaolong Wang*, Hao Su*

[Paper] [Website]

Method

TD-MPC is a framework for model predictive control (MPC) using a Task-Oriented Latent Dynamics (TOLD) model and a terminal value function learned jointly by temporal difference (TD) learning. TD-MPC plans actions entirely in latent space using the TOLD model, which learns compact task-centric representations from either state or image inputs. TD-MPC solves challenging Humanoid and Dog locomotion tasks in 1M environment steps.

Citation

If you use our method or code in your research, please consider citing the paper as follows:

@article{Hansen2022tdmpc,
	title={Temporal Difference Learning for Model Predictive Control},
	author={Nicklas Hansen and Xiaolong Wang and Hao Su},
	eprint={2203.04955},
	archivePrefix={arXiv},
	primaryClass={cs.LG},
	year={2022}
}

Instructions

Assuming that you already have MuJoCo installed, install dependencies using conda:

conda env create -f environment.yaml
conda activate tdmpc

After installing dependencies, you can train an agent by calling

python src/train.py task=dog-run

Evaluation videos and model weights can be saved with arguments save_video=True and save_model=True. Refer to the cfgs directory for a full list of options and default hyperparameters, and see tasks.txt for a list of supported tasks. We also provide results for all 23 state-based DMControl tasks in the results directory.

The training script supports both local logging as well as cloud-based logging with Weights & Biases. To use W&B, provide a key by setting the environment variable WANDB_API_KEY=<YOUR_KEY> and add your W&B project and entity details to cfgs/default.yaml.

Changelog

• [08-29-2022] Added safeguard against NaNs in rare cases. Fixed an issue that caused multi-dimensional observation spaces to be inferred incorrectly.
• [03-27-2022] Reduced memory usage in pixel experiments by 6x. Code improvements. Refactoring. Update default pixel hyperparameters.
• [03-10-2022] Initial code release.

License & Acknowledgements

TD-MPC is licensed under the MIT license. MuJoCo and DeepMind Control Suite are licensed under the Apache 2.0 license. We thank the DrQv2 authors for their implementation of DMControl wrappers.