This is a library for solving global trajectory optimization problems given problem specific derived classes for:
1) A dynamical model for the system (i.e. x'(t)=f(x(t),u(t)$)
2) A cost function of a trajectory $x(t)$ and input signal u(t) to be minimized ( i.e. C(x,u)= Integral g(x(t),u(t)) )
3) A collision detector. That is, if some subset of the state space must be avoided, then function must be provided to determine if a trajectory intersects that region
4) An admissible heuristic to be used in a graph search. Note that the heuristic h(x)=0 is always admissible.
This library may be considered as an alternative to a) Sampling-based planners b) Variational trajectory optimization utilizing nonlinear programming techniques
This method handles differential constraints that arise in nonholonomic planning and planning for complex dynamical systems. The reason is that optimal sampling-based planners (such as RRT*) require a steering subroutine in addition to (1)-(4) above which is non-trivial to provide in general. The downside is that for holonomic planning, such as simple shortest path queries, RRT* and PRM* achieve better performance in comparable implementations.
This method is far more robust than nonlinear programming based techniues. It does not require a "good initial guess" and it will not converge to a locally optimal solution. The downside is the complexity of this method is exponential with the state space. Up to 3 dimensional state spaces are dealt with very well. In higher dimensions, some domain knowledge will be required to construct a good heuristic.
A technical paper describing the theoretical aspects of the method can be found here:
https://arxiv.org/abs/1607.06966
Documentation of the C++ implementation can be found at the link below or by running doxygen from the top level directory:
https://codedocs.xyz/bapaden/Global-Trajectory-Optimization/md_README.html
To install the library after downloading the source code, enter the following terminal commands from the top level directory
mkdir build
cd build
cmake ..
make
make test
sudo make install
The unit tests are run using GTest. If it is not installed in your image you can install it with apt, or compile from source
sudo apt-get install libgtest-dev
https://github.com/google/googletest
Several basic examples demonstrating how to interface with the library can ve found in the examples/ directory. Each example generates data that is saved in the plots/ directory which contains python scripts to generate basic illustrations of the solution. To run the examples:
cd GlobalTrajectoryOptimization/build/examples
./shortest-path-demo
./pendulum-swingup-demo
./nonholonomic-car-demo
To view the solutions:
cd GlobalTrajectoryOptimization/examples
python shortest_path_viewer.py
Using Cmake, you can link to the installed library as follows:
find_package(glc)
include_directories(${GLC_INCLUDE_DIRS})
add_executable(your_awesome_planning_algorithm your_src.cpp)
target_link_libraries(your_awesome_planning_algorithm glc_planner_core)
In your source code where you instantiate a Planner object include the header
#include<glc/glc_planner_core.h>
You will have to implement derived classes for the following virtual base classes: (a) DynamicalSystem, (b) CostFunction, (c) Heuristic, (d) GoalRegion, (e) Obstacle. These base classes will need to meet the requirements described in the technical paper as well as the source documentation.