Airsim 配置与自定义说明
1. 配置说明
推荐使用双电脑仿真，一台windos、一台ubuntu。这样做的好处是可以方便配置，并且保证有很好的仿真性能。具体的Airsim配置，不管是官网还是其他博客都有很好的说明，这里我不再赘述，只把相应的链接放在这里，大家照着做就好。
- windows 配置
官网说明：https://microsoft.github.io/AirSim/build_windows/
更为细致的个人博客：https://www.zhihu.com/column/multiUAV （注意，这篇博客是Airsim 1.5版本。如果要配置最新版本请再参考一下英文官网（中文的wiki说明也是旧版本的）
在上述教程中提供了自定义无人机的外观的例子。不过需要的注意的是，它只是改变了渲染外形，并不改变实际的气动模型和质量，转动惯量等物理特性。另外根据我个人的使用经验，后续我们在加入自己的无人机时，推荐采用solidwors绘制模型，blender转换格式的方法来实现我们自己的无人机外形。
通过上述的材料，如果不出意外，你已经基本上手了Airsim的使用（在windows下）。
- 传感器添加参考（一定要加，不然ROS里就没有传感器数据）
{
"SettingsVersion": 1.2,
"SimMode": "Multirotor",
"DefaultVehicleConfig": "SimpleFlight",
"ViewMode": "FlyWithMe",
"UsageScenario": "",
"LocalHostIp": "192.168.3.99",
"RpcEnabled": true,
"Recording": {
"RecordOnMove": false,
"RecordInterval": 0.05,
"Cameras": [
{
"CameraID": 0,
"ImageType": 0,
"PixelsAsFloat": false,
"Compress": true
}
]
},
"SubWindows": [
{
"WindowID": 0,
"CameraName": "front_left",
"ImageType": 0,
"Visible": true
}
],
"CameraDefaults": {
"CaptureSettings": [
{
"ImageType": 0,
"Width": 640,
"Height": 480,
"FOV_Degrees": 90,
"AutoExposureSpeed": 100,
"AutoExposureBias": 0,
"AutoExposureMaxBrightness": 0.64,
"AutoExposureMinBrightness": 0.03,
"MotionBlurAmount": 0,
"TargetGamma": 1.0
}
],
"NoiseSettings": [
{
"ImageType": 0,
"Enabled": true,
"RandContrib": 0.2,
"RandSpeed": 100000.0,
"RandSize": 1000.0,
"RandDensity": 6.5,
"HorzWaveContrib": 0,
"HorzWaveStrength": 0.08,
"HorzWaveVertSize": 1.0,
"HorzWaveScreenSize": 1.0,
"HorzNoiseLinesContrib": 0,
"HorzNoiseLinesDensityY": 0.01,
"HorzNoiseLinesDensityXY": 0.5,
"HorzDistortionContrib": 0,
"HorzDistortionStrength": 0.002
}
]
},
"Vehicles": {
"drone_1": {
"VehicleType": "SimpleFlight",
"DefaultVehicleState": "Armed",
"EnableCollisionPassthrogh": false,
"EnableCollisions": true,
"AllowAPIAlways": true,
"RC": {
"RemoteControlID": 0,
"AllowAPIWhenDisconnected": false
},
"Cameras": {
"front_left": {
"CaptureSettings": [
{
"ImageType": 0,
"Width": 640,
"Height": 480,
"FOV_Degrees": 90,
"TargetGamma": 1.0
}
],
"X": 0.52,
"Y": -0.095,
"Z": 0,
"Pitch": 0,
"Roll": 0,
"Yaw": 0
},
"front_right": {
"CaptureSettings": [
{
"ImageType": 0,
"Width": 640,
"Height": 480,
"FOV_Degrees": 90,
"TargetGamma": 1.0
}
],
"X": 0.52,
"Y": 0.095,
"Z": 0,
"Pitch": 0,
"Roll": 0,
"Yaw": 0
},
"bottom_center": {
"CaptureSettings": [
{
"ImageType": 0,
"Width": 640,
"Height": 480,
"FOV_Degrees": 90,
"TargetGamma": 1.0
}
],
"X": 0.0,
"Y": 0.0,
"Z": 0.1,
"Pitch": -90,
"Roll": 0,
"Yaw": 0
}
}
}
}
}
- Ubuntu 配置
为了方便windows和Ubuntu通信，请将有线连接的ip设置为一个你喜欢的值（eg: 192.168.3.99)，还要记得关闭windows的防火墙，不然ping不通。
接下来，简要说明一下ROS和Airsim的配合使用。下面给出的官方说明的是windows下WSL的使用，这里我们使用ubuntu是和WSL1一样的操作方法，只要写好windows的ip即可。当然要记得先配置ubuntu的有线网络连接，保证和windows在同一网段（eg: 192.168.3.9)。
How to run Airsim on Windows and ROS wrapper on WSL#
For WSL 1 execute: export WSL_HOST_IP=127.0.0.1 and for WSL 2: export WSL_HOST_IP=$(cat /etc/resolv.conf | grep nameserver | awk '{print $2}') Now, as in the running section for linux, execute the following:
source devel/setup.bash
roslaunch airsim_ros_pkgs airsim_node.launch output:=screen host:=$WSL_HOST_IP
roslaunch airsim_ros_pkgs rviz.launch
2.自定义说明
为了能够更加逼近实际的飞行，做出更准确的飞行控制，实现从仿真到实机的无缝切换，只用Airsim自带的无人机模型大概是不够的。从Airsim官方文献 ，我们可以看到，其所建立的旋翼气动模型其实是很粗糙的，其精度远远不满足科研级别的要求，如果我们需要做无人机的高速飞行敏捷控制，这从仿真移植到实机将会是致命的灾难。
注意 ：输入控制信号和实际输出之间有一个一阶滤波器（可以充当惯性环节的电机响应，文献中没有写）。
static void setOutput(Output& output, const RotorParams& params, const FirstOrderFilter<real_T>& control_signal_filter, RotorTurningDirection turning_direction)
{
output.control_signal_input = control_signal_filter.getInput();
output.control_signal_filtered = control_signal_filter.getOutput();
//see relationship of rotation speed with thrust: http://physics.stackexchange.com/a/32013/14061
output.speed = sqrt(output.control_signal_filtered * params.max_speed_square);
output.thrust = output.control_signal_filtered * params.max_thrust;
output.torque_scaler = output.control_signal_filtered * params.max_torque * static_cast<int>(turning_direction);
output.turning_direction = turning_direction;
}
- 动力学
不过，从Airsim的代码框架中，要对旋翼模型做出修改并不难。在/Airlib/include/vehicles/multirotor/RotorActuator.hpp和/Airlib/include/vehicles/multirotor/RotorParams.hpp中修改源代码即可。另外飞行器阻力也应该进行修改（搜索一下drag)
此外，除了旋翼模型，我们也需要和真机质量和转动惯量更相符合的刚体模型。在/Airlib/include/vehicles/multirotor/MultiRotorParams.hpp中void setupFrameGenericQuad(Params& params)修改即可。
最后，如果需要其它的无人机动力学仿真，那么可以参考这个issue
更进一步阻力模型的建立： 有一说一，这个阻力模型也挺粗糙的。（在这里我们可以注意到如何将风添加到动力学模型中，这对我们修改旋翼气动模型也是有用的）
// drag term in Airsim
/**\AirSim\AirLib\include\vehicles\multirotor\MultiRotorPhysicsBody.hpp**/
void createDragVertices()
{
const auto& params = params_->getParams();
//Drone is seen as central body that is connected to propellers via arm. We approximate central body as box of size x, y, z.
//The drag depends on area exposed so we also add area of propellers to approximate drag they may introduce due to their area.
//while moving along any axis, we find area that will be exposed in that direction
real_T propeller_area = M_PIf * params.rotor_params.propeller_diameter * params.rotor_params.propeller_diameter;
real_T propeller_xsection = M_PIf * params.rotor_params.propeller_diameter * params.rotor_params.propeller_height;
real_T top_bottom_area = params.body_box.x() * params.body_box.y();
real_T left_right_area = params.body_box.x() * params.body_box.z();
real_T front_back_area = params.body_box.y() * params.body_box.z();
Vector3r drag_factor_unit = Vector3r(
front_back_area + rotors_.size() * propeller_xsection,
left_right_area + rotors_.size() * propeller_xsection,
top_bottom_area + rotors_.size() * propeller_area)
* params.linear_drag_coefficient / 2;
//add six drag vertices representing 6 sides
drag_faces_.clear();
drag_faces_.emplace_back(Vector3r(0, 0, -params.body_box.z() / 2.0f), Vector3r(0, 0, -1), drag_factor_unit.z());
drag_faces_.emplace_back(Vector3r(0, 0, params.body_box.z() / 2.0f), Vector3r(0, 0, 1), drag_factor_unit.z());
drag_faces_.emplace_back(Vector3r(0, -params.body_box.y() / 2.0f, 0), Vector3r(0, -1, 0), drag_factor_unit.y());
drag_faces_.emplace_back(Vector3r(0, params.body_box.y() / 2.0f, 0), Vector3r(0, 1, 0), drag_factor_unit.y());
drag_faces_.emplace_back(Vector3r(-params.body_box.x() / 2.0f, 0, 0), Vector3r(-1, 0, 0), drag_factor_unit.x());
drag_faces_.emplace_back(Vector3r( params.body_box.x() / 2.0f, 0, 0), Vector3r( 1, 0, 0), drag_factor_unit.x());
}
/** \AirSim\AirLib\include\physics\FastPhysicsEngine.hpp **/
static Wrench getDragWrench(const PhysicsBody& body, const Quaternionr& orientation,
const Vector3r& linear_vel, const Vector3r& angular_vel_body, const Vector3r& wind_world)
{
//add linear drag due to velocity we had since last dt seconds + wind
//drag vector magnitude is proportional to v^2, direction opposite of velocity
//total drag is b*v + c*v*v but we ignore the first term as b << c (pg 44, Classical Mechanics, John Taylor)
//To find the drag force, we find the magnitude in the body frame and unit vector direction in world frame
//http://physics.stackexchange.com/questions/304742/angular-drag-on-body
//similarly calculate angular drag
//note that angular velocity, acceleration, torque are already in body frame
Wrench wrench = Wrench::zero();
const real_T air_density = body.getEnvironment().getState().air_density;
// Use relative velocity of the body wrt wind
const Vector3r relative_vel = linear_vel - wind_world;
const Vector3r linear_vel_body = VectorMath::transformToBodyFrame(relative_vel, orientation);
for (uint vi = 0; vi < body.dragVertexCount(); ++vi) {
const auto& vertex = body.getDragVertex(vi);
const Vector3r vel_vertex = linear_vel_body + angular_vel_body.cross(vertex.getPosition());
const real_T vel_comp = vertex.getNormal().dot(vel_vertex);
//if vel_comp is -ve then we cull the face. If velocity too low then drag is not generated
if (vel_comp > kDragMinVelocity) {
const Vector3r drag_force = vertex.getNormal() * (- vertex.getDragFactor() * air_density * vel_comp * vel_comp); //this is wrong! TLK 2022-9-14
const Vector3r drag_torque = vertex.getPosition().cross(drag_force);// drag_torque is always zero
wrench.force += drag_force;
wrench.torque += drag_torque;
}
}
- 天气
https://zhaoxuhui.top/blog/2021/12/06/airsim-note7-weather-system-in-airsim.html
https://zhuanlan.zhihu.com/p/425804207
总结
Airsim 作为微软推出的用于推动自动驾驶和飞行的一款高保真仿真软件，其整个框架和文档是清晰完善的。基于unreal engine, Airsim可以得到保真度很高的视频画面，用于端到端的AI训练与部署。但是Airsim中无人机的动力学引擎部分依旧是粗糙的，对于低速场景其可以使用。但是对于高速飞行来收，保真度与state of the art的文章来说，有较大差距。如果后续对控制规划做深入深入，这部分需要参考相应的文献来进行修改。
1382
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
