STORM

Stochastic Tensor Optimization for Robot Motion - A GPU Robot Motion Toolkit

[Install Instructions] [Paper] [Website]

This package contains code for reactive robot motion leveraging parallel compute on the GPU. The implemented control framework leverages MPPI to optimize over sampled actions and their costs. The costs are computed by rolling out the forward model from the current state with the sampled actions. Most files are documented with sphinx. Once you clone this repo, go into docs folder and run sh generate_docs.sh to generate documentation.

To run on a real Franka Panda, you can use this low-level control wrapper: franka_motion_control from Mohak Bhardwaj.

Updates

Jan. 2022 - Add CoRL citation, merge torch.size() bug (thanks @maxpahn).

Contributors

• Mohak Bhardwaj
• Balakumar Sundaralingam

Citation

If you use this source code, please cite the below article,

@article{storm2021,
  title={{STORM}: An Integrated Framework for Fast Joint-Space Model-Predictive Control for Reactive Manipulation},
  author={Mohak Bhardwaj and Balakumar Sundaralingam and Arsalan Mousavian and Nathan D. Ratliff and Dieter Fox and Fabio Ramos and Byron Boots},
  booktitle={5th Annual Conference on Robot Learning },
  year={2021},}

Contributing to this code

Refer to CLA before making contributions.

need to export env variable before use(here replace the andylee with your user name)

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/andylee/miniconda3/envs/storm_kit/lib

python isaac sim trajectory follower demo

python3 examples/franka_reacher_traj.py

ros conda enrionment and usage, for integration with global planner

first you should install ros noetic on your ubuntu system then do following cmd in the conda environment

source /opt/ros/noetic/setup.bash
pip install rospkg

then you can run the ros_storm as:

python3 examples/franka_reacher_traj_ros.py