Original PyTorch implementation of MoDem-V2: Visuo-Motor World Models for Real-World Robot Manipulation by
Patrick Lancaster, Nicklas Hansen, Aravind Rajeswaran Vikash Kumar (Meta AI, UC San Diego)
MoDem-V2 combines the sample efficiency of the original MoDem with conservative exploration in order to quickly and safely learn manipulation skills on real robots.
If you use this repo in your research, please consider citing the paper as follows:
@article{lancaster2023modem,
title={MoDem-V2: Visuo-Motor World Models for Real-World Robot Manipulation},
author={Lancaster, Patrick and Hansen, Nicklas and Rajeswaran, Aravind and Kumar, Vikash},
journal={arXiv preprint arXiv:2309.14236},
year={2023}
}
We assume that your machine has a CUDA-enabled GPU, a local copy of MuJoCo 2.1.x installed, and at least 80GB of memory. Then, create a conda environment with conda env create -f environment.yml, and add modemv2/tasks/robohive to your PYTHONPATH. Activate the new environment with conda activate modemv2 and then install mujoco_py with pip install -e ./mujoco_py. You will also need to configure wandb_entity in modemv2/cfgs/config.yaml. Demonstrations are made available here; untar them into modemv2/demonstrations.
Launch MoDem-V2 training with the scripts in scripts/franka. Note that the scripts should be executed from the modemv2 directory. For example, to train a single seed of MoDem-V2 on the bin picking task:
sh scripts/franka/bin_pick/modemv2.sh
Append an argument of 1 in order to train 5 seeds on the cluster, for example:
sh scripts/franka/bin_pick/modemv2.sh 1
Alternatively, append an argument of 2 in order to truncate each stage of training and verify that the code has been setup correctly, for example:
sh scripts/franka/bin_pick/modemv2.sh 2
This codebase is based on the original MoDem implementation. The majority of MoDem-V2 is licensed under CC-BY-NC, however portions of the project are available under separate license terms: mujoco-py is licensed under the following license: https://github.com/openai/mujoco-py/blob/master/LICENSE.md; robohive is licensed under the following license: https://github.com/vikashplus/robohive/blob/main/LICENSE.