[Website] [PDF] [Arxiv]

This repository contains the code for the paper "Model-based Diffusion for Trajectory Optimization".

Model-based diffusion (MBD) is a novel diffusion-based trajectory optimization framework that employs a dynamics model to approximate the score function. MBD outperforms existing methods (including RL) in terms of sample efficiency and generalization.

To install the required packages, run the following command:

git clone --depth 1 git@github.com:LeCAR-Lab/model-based-diffusion.git
pip install -e .

To run model-based diffusion to optimize a trajectory, run the following command:

cd mbd/planners
python mbd_planner.py --env_name $ENV_NAME

where $ENV_NAME is the name of the environment, you can choose from hopper, halfcheetah, walker2d, ant, humanoidrun, humanoidstandup, humanoidtrack, car2d, pushT.

To run model-based diffusion combined with demonstrations, run the following command:

cd mbd/planners
python mbd_planner.py --env_name $ENV_NAME --enable_demos

Currently, only the humanoidtrack, car2d support demonstrations.

To run multiple seeds, run the following command:

cd mbd/scripts
python run_mbd.py --env_name $ENV_NAME

To visualize the diffusion process, run the following command:

cd mbd/scripts
python vis_diffusion.py --env_name $ENV_NAME

Please make sure you have run the planner first to generate the data.

To run model-based diffusion for black-box optimization, run the following command:

cd mbd/blackbox
python mbd_opt.py

To run RL-based baselines, run the following command:

cd mbd/rl
python train_brax.py --env_name $ENV_NAME

To run other zeroth order trajectory optimization baselines, run the following command:

cd mbd/planners
python path_integral.py --env_name $ENV_NAME --mode $MODE

where $MODE is the mode of the planner, you can choose from mppi, cem, cma-es.

• This codebase's environment and RL implementation is built on top of Brax.

@misc{pan2024modelbaseddiffusiontrajectoryoptimization,
      title={Model-Based Diffusion for Trajectory Optimization}, 
      author={Chaoyi Pan and Zeji Yi and Guanya Shi and Guannan Qu},
      year={2024},
      eprint={2407.01573},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2407.01573}, 
}