Motion Planning Networks
Implementation of MPNet: Motion Planning Networks. [arXiv1] [arXiv2]
The code can easily be adapted for Informed Neural Sampling.
Contains
- Data Generation
- Any existing classical motion planner can be used to generate datasets. However, we provide following implementations in C++:
- MPNet algorithm
- A navie python visualization files
Data Description
- Simple 2D has 7 blocks each of size 5x5 that are placed randomly.
- Complex 3D contains 10 blocks with sizes as follow:
- shape=[[5.0,5.0,10.0],[5.0,10.0,5.0],[5.0,10.0,10.0],
[10.0,5.0,5.0],[10.0,5.0,10.0],[10.0,10.0,5.0],
[10.0,10.0,10.0],[5.0,5.0,5.0],[10.0,10.0,10.0],[5.0,5.0,5.0]]
- e0-109 has the training and testing paths in 110 different environments.
- 0-100 environments and 0-4000 paths/environment are for training.
- Seen test dataset: 0-100 envs and 4000-4200=200 paths/env.
- Unseen test dataset: 100-110 envs and 0-2000 paths/env.
- obs_cloud is the point-cloud of randomly generated 30,000 environments.
- 0-110 corresponds to the same environments for which path data is provided.
- You may use full dataset to train encoder network via unsupervised learning.
- obs.dat contains the center location (x,y) of each obstacle in the environments.
- obs_perm2.dat contains the order in which the blocks should be placed in preset locations given by obs.dat file to setup environments.
- For instance, in complex 3D, the permutation 8342567901 indicates obstacle #8 of size 10x10x10 should be placed at the location #0 given by obs.dat.
Generating your own data
- Define a region of operation, for instance in simple2D, it is 20x20
- Decide how many obstacles (r) you would like to place in the region. In the case of simple2D, we have r=7 5x5 blocks.
- Generate random N locations to place r obstacles in the region. In the case of simple2D, we generated N=20.
- For N locations and r obstacles, apply combinatorics, to generate NCr different environments i.e., in simple 2D NCr= 20C7= 77520
- The obs_perm2 file contains the combinations, for instance 6432150 indicates to place obstacle#6 at location #0.
- Once obstacles are placed, randomly generate collision-free samples and use them in pairs as stat-goal to generate paths using any classical planne for the training. For classical planners, we recommend using OMPL implementations.
Requirements
-
Data Generation
-
Install libbot2
- Make sure all dependencies of libbot2 (e.g., lcm) are installed.
- Install libbot2 with the local installation procedure.
- Run "make" in the data_generation folder where the README file is located.
-
Use any compiler such as Netbeans to load the precomplie code.
-
data_generation/src/rrts_main.cpp contains the main rrt/prrt code.
-
data_generation/viewer/src/viewer_main.cpp contains the visualization code.
- Also checkout comments in data_generation/viewer/src/renderers/graph_renderer.cpp
-
Note: main_viewer and rrts_main should run in parallel as:
- rrts_main sends the path solution as well as the tree to the main_viewer to publish through local network.
- data is transmitted through LCM network protocol.
-
MPNet
Examples
-
Assuming paths to obstacles point-cloud are declared, train obstacle-encoder:
python MPNET/AE/CAE.py
-
Assuming paths to demonstration dataset and obstacle-encoder are declared, run mpnet_trainer:
python MPNET/train.py
-
Run tests by first loading the trained models:
python MPNET/neuralplanner.py
References
@inproceedings{qureshi2019motion,
title={Motion planning networks},
author={Qureshi, Ahmed H and Simeonov, Anthony and Bency, Mayur J and Yip, Michael C},
booktitle={2019 International Conference on Robotics and Automation (ICRA)},
pages={2118--2124},
year={2019},
organization={IEEE}
}
@inproceedings{qureshi2018deeply,
title={Deeply Informed Neural Sampling for Robot Motion Planning},
author={Qureshi, Ahmed H and Yip, Michael C},
booktitle={2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
pages={6582--6588},
year={2018},
organization={IEEE}
}
@article{qureshi2019motion,
title={Motion Planning Networks: Bridging the Gap Between Learning-based and Classical Motion Planners},
author={Qureshi, Ahmed H and Miao, Yinglong and Simeonov, Anthony and Yip, Michael C},
journal={arXiv preprint arXiv:1907.06013},
year={2019}
}