Attention: This work is based on unreal_cv_ros from ethz-asl. Because the original work is developed on UE4.16, a relatively old version of UE. I tuned some code to make it work on UE4.27.2.
The UnrealCV branch I choose is 4.27-stable. You need add some code just like original unreal_cv_ros does. The code is in the CustomPluginCode.cpp
file. You can find it in the unreal_cv_ros/content
folder.
Below is the original README.md file.
unreal_cv_ros is a package to allow ROS based simulation of a MAV equipped with a 3D-reconstruction sensor. The simulation is performed inside an Unreal Engine 4 (UE4) game. The node-game communcation is carried out utilizing the UnrealCV computer vision plugin for UE4.
- Note: We recommend also checking out our recent robot simulator https://github.com/ethz-asl/unreal_airsim for UE4 based on Microsoft AirSim with more features and higher performance.
Installation
Examples
Documentation: ROS Nodes
Documentation: Working with Unreal
- UnrealCV Plugin Setup
- Creating UE4 worlds
- When to use which mode
- Pawns and Cameras
- Custom collision and camera settings
- Static mesh collision
- Producing ground truth pointclouds
- The Unreal Coordinate System
Troubleshooting
System Dependencies:
Unreal_cv_ros requires the UnrealCV python library: pip install unrealcv
.
ROS Packages:
The perception functionalities (unreal_ros_client + sensor_model) depend on:
catkin_simple
(https://github.com/catkin/catkin_simple)
To run the full MAV simulation, these additional packages are needed:
rotors_simulator
(https://github.com/ethz-asl/rotors_simulator)mav_control_rw
(https://github.com/ethz-asl/mav_control_rw)voxblox
(https://github.com/ethz-asl/voxblox)
Install the ROS package:
Installation instructions on Linux:
- Move to your catkin workspace:
cd ~/catkin_ws/src
- Install using a SSH key or via HTTPS:
git clone [email protected]:ethz-asl/unreal_cv_ros.git # SSH
git clone https://github.com/ethz-asl/unreal_cv_ros.git # HTTPS
- Compile:
catkin build unreal_cv_ros
Install Unreal Engine:
To create custom worlds or run non-shipping projects, the Unreal Engine Editor is used. On Linux, this requires installing Unreal Engine and building it from source. Complete installation instrcutions are given on their webpage. In addition, install the UnrealCV plugin to Unreal Engine as described in UnrealCV Plugin Setup.
The unreal_cv_ros pipeline also works with stand-alone games that were built including the UnrealCV plugin (e.g. the examples below), which do not require additional installations.
Other related resources, such as blueprint classes and experiment scenarios, can be downloaded from here. This repo and the related resources were developped and tested with unrealcv v0.3.10 and Unreal Engine 4.16.3.
To illustrate the vision pipeline in stand-alone fashion, we run the unreal_ros_client in test mode with ground_truth as our sensor model.
View of the realistic rendering demo (left) and the produced ground truth point cloud with corresponding camera pose (right).
Please download the RealisticRendering game binary and launch the game. In a command window type roslaunch unreal_cv_ros example_test.launch
to start the pipeline and wait until the connection is setup (takes few seconds). You can now navigate the camera inside the game using the mouse and W-A-S-D keys while a rviz window displayes the produced ground truth pointclouds as well as the camera pose in the unreal world frame.
Note that the test mode is mostly meant to explore the effects of different sensor models on the pointcloud. Since the taking of images and read-out of the unreal-pose is done sequentially, fast movement in the game may result in some frames not being well aligned.
This example demonstrates how to setup the full scale MAV simulation. It uses unreal_cv_ros for perception and collision checking, gazebo to model the MAV physics, a MPC high level and PID low level controller for trajectory tracking and voxblox to integrate the pointclouds into a map.
Voxblox map that is continually built as the MAV follows a trajectory through the room.
Run the RealisticRendering demo as in the previous example. Make sure to tab out of it immediately to disable player control. This is necessary because unreal_cv_ros sets its world frame at the current MAV position on startup and this example expects the game demo to be in its initial state. Furthermore, if captured, unreal engine continues to accept player input that may interfere with the simulation.
In a command window type roslaunch unreal_cv_ros example_mav.launch
to start the pipeline. The simulation_manager will supervise the startup of all elements (this may take few seconds). After everything set up cleanly, the example node will publish two trajectory segments, which the MAV tries to follow to explore the room. In a rviz window, the current MAV pose is depicted together the planned trajectories and the voxblox mesh representation of the room as it is explored.
This node manages the unrealcv client and the connection with a running UE4 game. It sets the camera position and orientation within the unreal game and produces the raw image data and camera calibration used by the sensor_model.
-
mode In which mode the client is operated. Currently implemented are:
- test Navigate the MAV manually in the unreal game. The client will periodically publish the sensor data.
- standard The camera pose is set based on the
/odometry
topic and images are taken. Uses the default unrealcv plugin, operates at ~1 up to 2 Hz. - fast Similar to standard, but requires the custom unrealcv plugin (See Unrealcv Plugin Setup). Operates at ~3 up to 5 Hz.
Default is 'standard'.
-
collision_on Set to true to check for collision in the unreal game. Set to false to set the camera anyway. May result in rendering artifacts if the camera overlaps with objects. Default is true.
-
collision_tol This parameter only shows in
standard
mode. Collision warnings are triggered if the requested and realized position are further apart than this threshold (in unreal units, default unit is cm). Default is 10. -
publish_tf If true, the client pulishes a tf-transform of the camera pose for every taken image with a matching timestamp. Default is False.
-
slowdown Artificially slows down the time between setting the pose and taking images to give unreal engine more time to render the new view. Slowdown is expected as wait duration in seconds wall-time. Default is 0.0.
-
camera_id Lets unrealcv know which camera to use. Default is 0.
-
queue_size Queue size of the input subscriber. Default is 1.
-
Camera Parameters: To change the resolution and field of view (FOV) of the camera, the unrealcv configuration file needs to be changed. The relevant path is displayed when the unreal_ros_client node is launched. When the client is setup, these values are published as 'camera_params' on the ros parameter server for other nodes to access them.
- odometry of type
nav_msgs.msg/Odometry
. Set the camera pose w.r.t. its position and yaw at the connection of the client. Does not appear intest
mode.
- ue_sensor_raw of type
unreal_cv_ros.msg/UeSensorRaw
. The output of the in-game image capture, containing a color and a depth image encoded as npy binaries. - collision of type
std_msgs.msg/String
. Publishes "MAV collision detected!" upon collision detection. Only available if collision_on is true.
- terminate_with_reset of type
std_srvs.srv/SetBool
. Stop processing new odom requests and reset the camera to the initial pose.
This node converts the unreal_ros_client output into a pointcloud for further processing. Sensor specific behaviour can be simulated by artificially altering the ground truth data.
-
model_type Which sensor to simulate. Currently implemented are:
- ground_truth: Produces the ground truth pointcloud without additional processing.
- kinect: Simulates a Kinect 3D sensor according to this paper. Notice that the sensor range is cropped to [0.5, 3.0] m and that the inclination angle is neglected, since it is constant up until ~60, 70 degrees.
- gaussian_depth_noise: Apply a gaussian, depth dependent term to the pointcloud z coordinate. Allows to set the params
k_mu_<i>
andk_sigma_<i>
for i in [0, 3], where f(z) = k0 + k1 * z + k2 * z^2 + k3 * z^3. Default for all k is 0.0.
Default is 'ground_truth'.
-
camera_params_ns Namespace where to read the unreal camera parameters from, which are expected as {height, width, focal_length}. Notice that the sensor_model waits until the camera params are set on the ros parameter server (e.g. from the unreal_ros_client). Default is 'unreal_ros_client/camera_params'.
-
maximum_distance All points whose original ray length is beyond maximum_distance are removed from the pointcloud. Set to 0 to keep all points. Default is 0.0.
-
flatten_distance Sets the ray length of every point whose ray length is larger than flatten_distance to flatten_distance. Set to 0 to keep all points unchanged. Default is 0.0.
-
publish_color_images In addition to the point clouds publish the perceived images, encoded as 4 channel RGBA. Default is False.
-
publish_gray_images If true publish a gray scale image with every pointcloud (e.g. for integration with rovio). Default is False.
- ue_sensor_raw of type
unreal_cv_ros.msg/UeSensorRaw
. Output of the unreal_ros_client that is to be further processed.
- ue_sensor_out of type
sensor_msgs.msg/PointCloud2
. Resulting pointcloud after applying the simulated sensing pipeline. - ue_color_image_out of type
sensor_msgs.msg/Image
. Color image of the current view. Only published if publish_color_images is true. - ue_gray_image_out of type
sensor_msgs.msg/Image
. Gray scale image of the current view. Only published if publish_gray_images is true.
This node is used to launch the full MAV simulation using gazebo as a physics engine and an unreal game for perception and collision modeling. It is used to coordinate simulation setup and monitor the unreal_cv_ros pipeline performance.
- ns_gazebo Namespace of gazebo, including the node name. Default is '/gazebo'.
- ns_mav Namespace of the MAV, which is expected to end with the actual MAV name. Default is '/firefly'.
- monitor Set to true to measure the unreal vision pipeline's performance. Default is False.
- horizon How many datapoints are kept and considered for the performance measurement. Only available if monitor is true. Default is 10.
- ue_raw_in of type
unreal_cv_ros.msg/UeSensorRaw
. Output of the unreal_ros_client for performance measurements. Only available if monitor is true. - ue_out_in of type
sensor_msgs.msg/PointCloud2
. Output of the sensor model for performance measurements. This topic needs to be matched to check for correct startup.
- simulation_ready of type
std_msgs.msg/String
. After successful start of the pipeline publishes "Simulation Ready" to let other nodes start or take over.
- display_monitor of type
std_srvs.srv/Empty
. Print the current monitoring measurements to console. Only available if monitor is true.
For the modes test
and standard
the default UnrealCV plugin is sufficient. Install it to the project or engine as suggested on their website.
This additional command is required to operate in fast
mode.
-
Compiled Plugin The compiled plugin can be downloaded from the data repository.
-
Compile it yourself To compile the plugin, first create an unreal engine development environment as is explained here. Then change the source code of the unrealcv plugin in your project (eg.
UnrealProjects/playground/Plugins/UnrealCV/Source/UnrealCV/Private/Commands
) to include the new command:- In
CameraHandler.h
add the command declaration - In
CameraHandler.cpp
add the command dispatcher and function body.
- In
These 3 code snippets can be copied from unreal_cv_ros/content/CustomPluginCode.cpp
. Afterwards build the project as explained in the unrealcv docs. The compiled plugin can now be copied to other projects or the engine for use.
This command take images and then request the passed position and orientation. This leaves time for the engine to finish rendering before the next images are requested.
- Syntax:
vget /uecvros/full X, Y, Z, p, y, r, collision, cameraID
. All argumentes, except the cameraID which is a uint, are floats. All arguments must be set. - Input:
- (X, Y, Z): New target position in unreal units.
- (p, y, r): New target orientation in degrees (and unreal coordinate system).
- collision: Set to 1 to check collision, -1 to ignore it.
- cameraID: ID of the camera to use in compliance with the unrealcv structure. Generally use 0.
- Output:
- In case of collision returns "Collision detected!" as unicode.
- Otherwise returns the stacked binary image data as string, where the first half representes the color image and the second half the depth image, both as npy arrays.
In order to easily create unreal_cv_ros compatible UE4 worlds:
- Install and use unreal engine editor 4.16 for compatibility with unrealcv. (Note: Newer versions should work too but without guarantees).
- Make sure the unrealcv plugin is installed and activated in the current project (In the Editor check: Edit > Plugins > Science > Unreal CV, see unrealcv docs).
- Set the player pawn to a flying spectator type with collision: World Settings > Game Mode > Selected GameMode > Default Pawn Class := UecvrosDefaultPawn. (If this is read-only just change to a custom GameMode.)
- Highly recommended: If using the
fast
mode, use theUecvrosDefaultPawn
. The blueprint class can be downloaded from the data repository. - Otherwise use the unreal engine built in
DefaultPawn
.
- Highly recommended: If using the
The unreal_cv_ros plugin is designed to work with all sorts of unreal worlds, however performance depends on how the world and the plugin is setup:
Plugin | Unreal World | Method to use |
---|---|---|
Custom | Custom | When the plugin and unreal world are modifiable, it is highly recommended to use the fast mode with the UecvrosDefaultPawn . This is not only the fastest combination, but it is also more stable and produces accurate and well-aligned pointclouds. |
Custom | Static | If the unrealworld cannot be modified (usually it uses the DefaultPawn ), it is still recommended to use the fast mode, since it runs considerably faster. However, the default camera may pose obstacles, such as motion blurr, that hinders alignment of the recorded images with the requested pose. Adjust the 'slowdown' parameter of the unreal_ros_client to give unreal enough time to properly process the desired images. |
Static | Any | If the plugin cannot be modified, use the standard mode and the built in DefaultPawn . |
test mode |
Any | If you want to run the test mode, use the built in DefaultPawn to allow for suitable user control. |
Unrealcv works by capturing pawns and cameras in the unreal game, to which it attaches. Pawns are the physical entities that move around and collide with things, whereas cameras typically provide the views. Usually, pawns are not placed in the world, thus upon game start a Pawn will spawn at PlayerStart that is then possessed by player control and unrealcv.
- Cameras in unrealcv are stored in chronological order. If you want to access different cameras use the 'camera_id' param of the unreal_ros_client.
- The
DefaultPawn
is incoporates standard flying player control and a default view. However that incorporates all default features such as motion blurr when moving (or teleporting!) the pawn. - The
UecvrosDefaultPawn
inherits from theDefaultPawn
. Furthermore, it has a camera attached to it that can be modified. However this disables the default player camera orientation control (which is also the one accessed with the standard unrealcv plugin).
Custom camera settings can similarly be changed by editing the camera component. Notice that unrealcv overwrites certain camera settings (such as the resolution and field of view) using the unrealcv configuration file.
The default collision for the DefaultPawn
and UecvrosDefaultPawn
is a sphere of 35cm radius. For custom collision and camera, you need to create your own pawn blueprint (which inherits from the respective base class). E.g. an 'easy' way to create a custom DefaultPawn
is as follows:
- In the Modes window, search for 'DefaultPawn' and create an instance (drag and drop into the game world).
- Select the pawn instance and click Blueprints > Convert selected actor to blueprint class...
- Save the new class, e.g. in the content/blueprints folder as myDefaultPawn.
- The class should now open in the blueprint editor, where its components can be edited.
- To change the radius select the 'CollisionComponent', and under Details > Shape > SphereRadius := myValue. Notice that too short radii can allow the camera to enter certain objects, creating graphical artifacts.
- Save the blueprint. Set the World Settings > Game Mode > Selected GameMode > Default Pawn Class := myDefaultPawn
When creating UE4 worlds, it is worth double checking the player collision with static mesh actors (the scene). By default, unreal produces convex hulls as simple collision objects and checks against simple collision. However, this may result in faulty collision detection. Collision can be changed to match the visible mesh as follows (may degrade performance):
- Right click your object in the World Outliner and select "Edit 'your object'".
- Set Collision > Collision Complexity := Use Complex Collision As Simple.
- Save and close the editing window.
Current collision can be visualized by changing the viewmode in the unreal editor from "Lit" to "Player Collision".
For simulation evaluation, ground truth meshes can be exported from the unreal editor and further processed using tools such as CloudCompare. A voxblox compatible ground truth pointcloud can be generated as follows:
- In the Unreal Engine Editor,
- Select the objects of interes in the World Outliner
- File > Export Selected...
- Save as *.obj file. (Currently no need to export the material too.)
- Open CloudCompare,
- File > Import > my_mesh_export.obj
- Use 'Edit > Multiply/Scale > 0.01' to compensate for unreal engine units (default is cm).
- Use 'Edit > Apply transformation' to place and rotate the object relative to the PlayerStart settings from unreal. (I.e. with the origin at the PlayerStart location, x pointing towards the PlayerStart orientation, y-left and z-up).
- Use 'Edit > Crop' to remove meshes outside the region of interest and unreachable points, such as the ground plane.
- Click 'Sample points on a mesh' to create a pointcloud. (Currently no need to generate normals or color.)
- 'File > Save' and save as a *.ply in ASCII format
- With a TextEditor,
- Open my_gt_pointcloud.ply and remove the "comment Author" and "obj_info" fields (lines 3 and 4, these create errors with pcl::plyreader).
For application with the unreal_ros_client, the coordinate transformations are already implemented so no need to worry. For development/debugging tasks: Unreal and unrealcv use the following coordinate system:
- Unreal World The default coordinate system is X-forward, Y-right, Z-up. Default units are cm.
- Rotation Direction Positive rotation directions around the unreal world coordinate axes are mathematically positive around the X and Y axes and negative around Z axis.
- Rotation parametrization The Unreal engine interface and therefore also the unrealcv commands parse rotations as pitch-yaw-roll (pay attention to the order). However, inside the engine rotations are performed as Euler-ZYX rotations (i.e. roll-pitch-yaw). Default units are degrees.
Known Issues and what to do about them:
- Error message "Error addressing the unrealcv client. Try restarting the game.":
- Make sure that only a single unreal game or editor is running. When using the editor, the unrealcv plugin is loaded already during startup (without the game itself running!). Since unrealcv per default connects to
localhost:9000
the server will be blocked if any other instance of it is running.
- Make sure that only a single unreal game or editor is running. When using the editor, the unrealcv plugin is loaded already during startup (without the game itself running!). Since unrealcv per default connects to
- The produced pointclouds are not well aligned with the requested pose / are smeared out:
- If running in standard mode, consider switching to fast mode which is not only faster but also more stable. (In standard mode every orientation change is carried out as rotation movement inducing rotation offsets (such as motion blurr when using the DefaultPawn camera), especially for high rotation rates).
- Give unreal engine more time for rendering and settling by adjusting the
slowdown
paramter of the unreal_ros_client. (This problem typically occurs for low rendering/frame rates in the unreal game or when using the DefaultPawn).
- I get strange collision warnings where there shouldn't be any:
- Make sure you are not using simple collision, which can greatly inflate collision bounds of objects. See here for more info.
- I get collision warnings for quickly executed trajectories:
- The unreal teleport action with collision checking sweeps (as far as I know) a linear path between every request. Try increasing the unreal_ros_client's update rate or slowing down simulated time.
- Installing UE4.16: "xlocale.h not found":
- xlocale is no longer needed and this bug is fixed in future versions of UE4, removing the include statements of xlocale should fix the issue.