UrbanFly: Uncertainty-Aware Planning for Navigation Amongst High-Rises with Monocular Visual-Inertial SLAM Maps
UrbanFly: an uncertainty-aware realtime planning framework for quadrotor navigation in urban high-rise environments. A core aspect of UrbanFly is its ability to robustly plan directly on the sparse point clouds generated by a Monocular Visual Inertial SLAM (VINS) backend. Through UrbanFly we present two trajectory optimizers, The first optimizer uses gradient-free cross-entropy method to compute trajectories that minimize collision probability and smoothness cost. Our second optimizer is a simplified version of the first and uses a sequential convex programming optimizer initialized based on probabilistic safety estimates on a set of randomly drawn trajectories.Empowered by the algorithmic innovation, UrbanFly outperforms competing baselines in metrics such as collision rate, trajectory length, etc., on a high fidelity AirSim simulator augmented with synthetic and real-world dataset scenes
Preprint: https://arxiv.org/pdf/2204.00865.pdf
Simulation Demonstration Video: Link
The planner evaluations were done on an 8-Core Intel Core i7-10870H processor, with 16GB RAM and 500GB HDD, running Ubunut 18.04 and ros melodic. we do recommend using powerfull setup.
A simulation server with NVIDIA1070 graphics card, AMD Ryzen 7 3800x 8-core processor× 16 CPU, and 64GB RAM and 1TB HDD was setup that hosted the simulation environments within the Unreal Engine.
ROS Melodic
Unreal
AirSim ROS Wrapper
PlaneRCNN
farneback3d
cd ~/catkin_ws/src/
git clone https://github.com/sudarshan-s-harithas/UrbanFly.git
In the UrbanFly Planar Mapping Module we provide two independent packages for out SquareStreet and the UrbanScene 3D environment.
Note: If the Mapping package is complied at once ROS would raise an error informing that there are multiple packages with the same name.
A possible solution would be to compile and test SquareStreet and UrbanScene 3D environment independently.
catkin_make -j8
source devel/setup.bash
The Unreal Engine with the drone and the simulation environment would run on a remote server and the mapping and planning modules would be implemented on the local computer. Start the simulation and run the below command to connect Airsim to the remote host. Also do-not forget to configure ROS Master slave .
roslaunch airsim_ros_pkgs airsim_node.launch host:=<ENTER_IP>
Clone VINS-Mono repo and run the below command to apply the changes made in VINS-Mono required for UrbanFly. These changes append unique feature ids to feature points in the point cloud topic.
git apply vins_mono_changes.patch
Then launch the vins estimator node using this command
roslaunch vins_estimator minihattan.launch
In Synthetic dataset (SquareStreet), ground truth masks are used. Whereas, for UrbanScene3D, Plance-RCNN is used to segment the planar facades.
In either case, mapper node can launched using the following command.
roslaunch rpvio_estimator minihattan.launch
TODO: Add mapping script
Additionally, for UrbanScene3D, ros-wrapper over PlaneRCNN should be lauched.
To run the simulation among real world like buildings, download UrbanScene3D dataset
SCP-MMD planner is implemented in Python in the folder SCP-MMD planner. It can be launched using the below command, after VINS-Mono and Mapping node is initialized.
python planner_node.py
The command given above publishes the path. Inorder to run the virtual drone in airsim, the below script should be run from another terminal. Note: Make sure that airsim python package is installed before running this command
python path_commander.py
Once the VINS is initilized and the planar map of the environment can be observed we are all set to start the CEM planner. Use the command given below to start the planner.
rosrun UrbanFlyCEMPlanner CEMPlanner
We created the Square Street Dataset which is in the shape of a square, having 47 buildings spread over 0.16 square kilometers. The environment is feature-rich and, therefore, can be easily used for benchmarking monocular vision-based perception and planning pipelines. The whole environment is modular and can be easily expanded or configured to meet custom requirements.
The Dataset and its usage instructions is given in this link