Motion Primitive Library is a search-based planner to compute dynamically feasible trajectories for a quadrotor flying in an obstacle-cluttered environment. Our approach searches for smooth, minimum-time trajectories by exploring the map using a set of short-duration motion primitives. The primitives are generated by solving an optimal control problem and induce a finite lattice discretization on the state space which can be explored using a graph-search algorithm. The proposed approach is able to generate resolution-complete (i.e., optimal in the discretized space), safe, dynamically feasibility trajectories efficiently by exploiting the explicit solution of a Linear Quadratic Minimum Time problem. It does not assume a hovering initial condition and, hence, is suitable for fast online re-planning while the robot is moving.
More details about the algorithm can be found in following publications:
- S. Liu, N. Atanasov, K. Mohta, and V. Kumar, "Search-based Motion Planning for Quadrotors using Linear Quadratic Minimum Time Control", IROS 2017
- S. Liu, K. Mohta, N. Atanasov, and V. Kumar, "Towards Search-based Motion Planning for Micro Aerial Vehicles", arxiv 2018
- Add iterative plan in MapPlanner
- Remove dependence on SDL library, using OpenCV for plotting
- Reformat the repo structure
- Add yaw primitive
- Add potential function to perturb trajectory
- Separate
ellipsoid_planner
tompl_ros
Eigen
: apt install libeigen3-devYAML-CPP
: apt install libyaml-cpp-devOpenCV
: apt install libopencv-dev
or simply run following commands:
$ sudo apt-get update
$ sudo apt install -y libeigen3-dev libyaml-cpp-dev libproj-dev libopencv-dev cmake
mkdir build && cd build && cmake .. && make -j4
$ mv motion_primitive_library ~/catkin_ws/src
$ cd ~/catkin_ws & catkin_make_isolated -DCMAKE_BUILD_TYPE=Release
Run following command in the build
folder for testing the executables:
$ make test
If everything works, you should see the results as:
Total Test time (real) = 4.22 sec
Running tests...
Test project /home/sikang/thesis_ws/src/packages/mpl_ros/motion_primitive_library/build
Start 1: test_traj_solver
1/6 Test #1: test_traj_solver ........................ Passed 0.00 sec
Start 2: test_planner_2d
2/6 Test #2: test_planner_2d ......................... Passed 0.92 sec
Start 3: test_planner_2d_prior_traj
3/6 Test #3: test_planner_2d_prior_traj .............. Passed 0.93 sec
Start 4: test_planner_2d_with_yaw
4/6 Test #4: test_planner_2d_with_yaw ................ Passed 0.96 sec
Start 5: test_distance_map_planner_2d
5/6 Test #5: test_distance_map_planner_2d ............ Passed 1.33 sec
Start 6: test_distance_map_planner_2d_with_yaw
6/6 Test #6: test_distance_map_planner_2d_with_yaw ... Passed 2.39 sec
100% tests passed, 0 tests failed out of 6
Total Test time (real) = 6.54 sec
To link this lib properly, add following in the CMakeLists.txt
find_package(motion_primitive_library REQUIRED)
include_directories(${MOTION_PRIMITIVE_LIBRARY_INCLUDE_DIRS})
...
add_executable(test_xxx src/test_xxx.cpp)
target_link_libraries(test_xxx ${MOTION_PRIMITIVE_LIBRARY_LIBRARIES})
To run the planner, three components are required to be set properly:
We use theclass Waypoint
for the start and goal. A Waypoint
contains coordinates of position, velocity, etc and the flag use_xxx
to indicate the control input.
An example for 2D planning is given as:
Waypoint2D start, goal; // Initialize start and goal as Waypoint2D
start.pos = Vec3f(2.5, -3.5);
start.use_pos = true;
start.use_vel = true;
start.use_acc = false;
start.use_jrk = false;
start.use_yaw = false;
goal.pos = Vec3f(35, 2.5);
goal.control = start.control;
The flag use_xxx
indicates the planner to plan in different control space. For example, the above example code sets the control in ACC
space. Eight options are provided by setting following flags:
~ | VEL | ACC | JRK | SNP | VEL&YAW | ACC&YAW | JRK&YAW | SNP&YAW |
---|---|---|---|---|---|---|---|---|
use_pos: |
true |
true |
true |
true |
true |
true |
true |
true |
use_vel: |
false |
true |
true |
true |
false |
true |
true |
true |
use_acc: |
false |
false |
true |
true |
false |
false |
true |
true |
use_jrk: |
false |
false |
false |
true |
false |
false |
false |
true |
use_yaw: |
false |
false |
false |
false |
true |
true |
true |
true |
In equal, one can also set the attribute control
of Waypoint
for the same purpose:
~ | VEL | ACC | JRK | SNP | VEL&YAW | ACC&YAW | JRK&YAW | SNP&YAW |
---|---|---|---|---|---|---|---|---|
control: |
Control::VEL |
Control::ACC |
Control::JRK |
Control::SNP |
Control::VELxYAW |
Control::ACCxYAW |
Control::JRKxYAW |
Control::SNPxYAW |
Any planner needs a collision checking function, there are several utils in this package to checking collision for obstacles in different representations.
In the most common environment where obstacles are represented as voxels, we use class MapUtil
which is a template class that adapts to 2D (OccMapUtil
) and 3D (VoxelMapUtil
).
An example for initializing a 2D collision checking OccMapUtil
is given as:
std::shared_ptr<MPL::OccMapUtil> map_util; // Declare as a shared pointer
map_util.reset(new MPL::OccMapUtil); // Initialize map_util
map_util->setMap(origin, dim, data, resolution); // Set the map information
...
planner->setMapUtil(map_util); // Set collision checking util
Here origin
, dim
, data
and resolution
are user input.
Our planner takes control input to generate primitives. User need to specify it before start planning.
An example for the control input U
for 2D planning is given as following, in this case, U
simply include 9 elements:
decimal_t u_max = 0.5;
vec_E<VecDf> U;
const decimal_t du = u_max / num;
for(decimal_t dx = -u_max; dx <= u_max; dx += du )
for(decimal_t dy = -u_max; dy <= u_max; dy += du )
U.push_back(Vec2f(dx, dy));
...
planner->setU(U); // Set control input
After setting up above 3 required components, a plan thread can be launched as:
std::unique_ptr<MPL::OccMapPlanner> planner(new MPL::OccMapPlanner(true)); // Declare a 2D planner with verbose
planner->setMapUtil(map_util); // Set collision checking util
planner->setU(U); // Set control input
planner->setDt(1.0); // Set dt for each primitive
bool valid = planner->plan(start, goal); // Plan from start to goal
After compiling by cmake
, run following command for test a 2D planning in a given map:
$ ./build/test_planner_2d ./data/corridor.yaml
You should see following messages if it works properly:
[MapPlanner] PLANNER VERBOSE ON
[PlannerBase] set v_max: 1.000000
[PlannerBase] set a_max: 1.000000
[PlannerBase] set dt: 1.000000
...
++++++++++++++++++ PLANNER +++++++++++++++
+ w: 10.00 +
+ dt: 1.00 +
+ ds: 0.0100 +
+ dv: 0.1000 +
+ da: 0.1000 +
+ dj: 0.1000 +
+ v_max: 1.00 +
+ a_max: 1.00 +
+ j_max: -1.00 +
+ U num: 9 +
+ tol_dis: 0.50 +
+ tol_vel: 0.00 +
+ tol_acc: 0.00 +
+ alpha: 0 +
+heur_ignore_dynamics: 1 +
++++++++++ PLANNER +++++++++++
Start from new node!
goalNode fval: 358.000000, g: 353.000000!
Expand [291] nodes!
...
Reached Goal !!!!!!
MPL Planner takes: 9.000000 ms
MPL Planner expanded states: 615
Total time T: 35.000000
Total J: J(VEL) = 36.750000, J(ACC) = 1.500000, J(JRK) = 0.000000, J(SNP) = 0.000000
The output image is saved in the current folder: (blue dots show the expended states, blue and cyan circles indicate start and goal).
Run following command for test a 2D planning, it first finds a trajector in low dimensional space (acceleration-driven), then it uses the planned trajectory to refine for a trajectory in high dimensional space (jerk-driven):
$ ./build/test_planner_2d_prior_traj ./data/corridor.yaml
In the following output image, the black curve is the prior trajectory:
In some cases, the robot needs to move forward within the FOV of the camera or
range sensor such that the yaw needs to be considered when planning. MapPlanner
handles the yaw constraint properly.
Following image shows the output of running:
$ ./build/test_planner_2d_with_yaw ./data/corridor.yaml
In practical case, the robot wants to stay away from obstacles even though the nominal trajectory is collision free. To add a soft constraint based on the distance towards obstacles, one technique is to use the artificial potential field (APF). In this examplem, we show how to perturb a nominal trajectory based on the search-based method with APFs:
$ ./build/test_distance_map_planner_2d ./data/corridor.yaml
In addition, to do the perturbation iteratively, run the other node:
$ ./build/test_distance_map_planner_2d_iterative ./data/corridor.yaml
In a more comprehensive case, when the robot has limited FOV and sensing range, plan the trajectory that considers safety and yaw constraint:
$ ./build/test_distance_map_planner_2d_with_yaw ./data/corridor.yaml
This example illustrate the mpl_traj_solver
which is a smoothing tool to derive a smoother trajectory.
An example of generating trajectory from a given path, without obstacles:
$ ./build/test_traj_solver
Here we generate three different trajectories using the same path and time allocation: the red one is minimum velocity trajectory, the green one is the minimum acceleration trajectory and the blue one is the minimum jerk trajectory.
For API, please refer to Doxygen.
The interface with ROS for visualization and implementation can be found in mpl_ros
.