CLAMP: Combined Learning from Demonstration and Motion Planning

CLAMP is an probabilistic inference based approach which learns a a trajectory prior from human demonstrations and uses factor graph optimization for efficient skill reproduction. The baseline capabilities of CLAMP are described in Towards robust skill generalization: Unifying learning from demonstration and motion planning (CoRL 2017). Extensions of CLAMP to learn from demonstrations provided in clutter as well as incrementally updating the trajectory prior are provided in Learning Generalizable Robot Skills from Demonstrations in Cluttered Environments (IROS 2018).

This library is an implementation of CLAMP. The core library is developed in C++ language, alongside a MATLAB frontend. Demo scripts are provided in MATLAB to reproduce some of the results in our publications.

CLAMP is being developed by M. Asif Rana and Mustafa Mukadam at the Georgia Tech Institute for Robotics & Intelligent Machines.

Prerequisites

• CMake >= 2.6 (Ubuntu: sudo apt-get install cmake), compilation configuration tool.
• Boost >= 1.50 (Ubuntu: sudo apt-get install libboost-all-dev), portable C++ source libraries.
• GTSAM >= 4.0 alpha, a C++ library that implement smoothing and mapping (SAM) framework in robotics and vision. Here we use factor graph implementations and inference/optimization tools provided by GTSAM.
• GPMP2, a C++ library that implements Gaussian Process Motion Planner 2.

(To use all the features, install the MATLAB toolboxes for GTSAM and GPMP2)

Compilation & Installation

In the library folder excute:

$ mkdir build
$ cd build
$ cmake ..
$ make check  # optonal, run unit tests
$ make install

Matlab Toolbox

A Matlab toolbox is provided to use our library in Matlab. To enable Matlab toolbox during compilation:

$ cmake -DCLAMP_BUILD_MATLAB_TOOLBOX:OPTION=ON -DGTSAM_TOOLBOX_INSTALL_PATH:PATH=/path/install/toolbox ..
$ make install

After you install the Matlab toolbox, don't forget to add /path/install/toolbox to your Matlab path.

Usage

The MATLAB toolbox contains multiple demo scripts described as follows,

1. mainBatchCLAMP.m - learns trajectory prior from an already acquired dataset, followed by skill generalization 1.
2. mainIncrementalCLAMP.m - updates the prior as dataset is aggregated, followed by skill generalization 2.
3. mainBatchCLAMP_ObstacleWeightedPrior.m - learns prior from demonstrations provided in the presence of clutter 2.

Citing

If you use CLAMP in an academic context, please cite the following publications:

[1] 
@inproceedings{rana2017towards,
  title={Towards robust skill generalization: Unifying learning from demonstration and motion planning},
  author={Rana, Muhammad Asif and Mukadam, Mustafa and Ahmadzadeh, S Reza and Chernova, Sonia and Boots, Byron},
  booktitle={Proceedings of the 2017 Conference on Robot Learning (CoRL)},
  year={2017}
}

[2]
@inproceedings{rana2018learning,
  title={Learning Generalizable Robot Skills from Demonstrations in Cluttered Environments},
  author={Rana, M Asif and Mukadam, Mustafa and Ahmadzadeh, S Reza and Chernova, Sonia and Boots, Byron},
  booktitle={2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  pages={4655--4660},
  year={2018},
  organization={IEEE}
}

Questions & Bug reporting

Please use Github issue tracker to report bugs. For other questions please contact Asif Rana or Mustafa Mukadam .

License

CLAMP is released under the BSD license, reproduced in the file LICENSE in this directory.