Updates: July 22, 2024 -- Full install, simulation testing, and real-world testing instructions/code available! Includes updates to setup files, launch_evaluation.bash, and to datashare files. July 25, 2024 -- Training instructions available, including details on using our provided dataset or gathering your own in simulation.
This is the official repository for the paper "Vision Transformers for End-to-End Vision-Based Quadrotor Obstacle Avoidance" by Bhattacharya, et al. (2024) from GRASP, Penn. Please note that you may find plenty of legacy and messy code in this research project's codebase.
We demonstrate that vision transformers (ViTs) can be used for end-to-end perception-based obstacle avoidance for quadrotors equipped with a depth camera. We train policies that predict linear velocity commands from depth images to avoid obstacles via behavior cloning from a privileged expert in a simple simulation environment, and show that ViT models combined with recurrence layers (LSTMs) outperform baseline methods based on other popular learning architectures. Deployed on a real quadrotor, our method achieves zero-shot dodging behavior at speeds reaching 7m/s and on multi-obstacle environments.
Note that if you'd only like to train models, and not test in simulation, you can skip straight to section Train.
If you'd like to start a new catkin workspace, then a typical workflow is (note that this code has only been tested with ROS Noetic and Ubuntu 20.04):
cd
mkdir -p catkin_ws/src
cd catkin_ws
catkin init
catkin config --extend /opt/ros/$ROS_DISTRO
catkin config --merge-devel
catkin config --cmake-args -DCMAKE_BUILD_TYPE=Release -DCMAKE_CXX_FLAGS=-fdiagnostics-color
Once inside your desired workspace, clone this repository (note, we rename it to vitfly):
cd ~/catkin_ws/src
git clone git@github.com:anish-bhattacharya/ViT-for-quadrotor-obstacle-avoidance.git vitfly
cd vitfly
In order to replicate Unity environments similar to those we use for training and testing, you will need to download environments.tar (1GB) from Datashare and extract it to the right location (below). We provide the medium-level spheres scene and a trees scene. Other obstacle environments are provided by ICRA 2022 DodgeDrone Competition.
tar -xvf <path/to/environments.tar> -C flightmare/flightpy/configs/vision
You will also need to download our Unity resources and binaries. Download flightrender.tar (450MB) from Datashare and then:
tar -xvf <path/to/flightrender.tar> -C flightmare/flightrender
Then, install the dependencies via a given script:
bash setup_ros.bash
cd ../..
catkin build
source devel/setup.bash
cd src/vitfly
Download pretrained_models.tar (50MB) from Datashare. This tarball includes pretrained models for ConvNet, LSTMnet, UNet, ViT, and ViT+LSTM (our best model).
tar -xvf <path/to/pretrained_models.tar> -C models
For testing on a medium-level spheres or trees environment, edit the file flightmare/flightpy/configs/vision/config.yaml line 2-3 as following:
level: "spheres_medium" # spheres
env_folder: "environment_<any int between 0-100>"
level: "trees" # trees
env_folder: "environment_<any int between 0-499>"
When running the simulation (the following section), you can set any number N of trials to run. To run trials on the same, specified environment index, set datagen: 1 and rollout: 0. To run sequentially different environment indices upon each trial, set datagen: 0 and rollout: 1. For the latter more, the environments prefixed with custom_ are used. These set the static obstacles as dynamic, so that they can be moved to new positions upon each trial, within the same Unity instance.
You may also change the unity: scene: 2 scene index according to those provided in the Unity binary. The available environments and their drone starting positions are found in flightmare/flightpy/configs/scene.yaml.
The launch_evaluation.bash script launches Flightmare and the trained model for depth-based flight when using vision mode. To run one trial, run:
bash launch_evaluation.bash 1 vision
Some details: Change 1 to any number of trials you'd like to run. If you look at the bash script, you'll see multiple python scripts being run. envtest/ros/evaluation_node.py counts crashes, starts and aborts trials, and prints other statistics to the console. envtest/ros/run_competition.py subscribes to input depth images and passes them to the corresponding functions (located in envtest/ros/user_code.py) that run the model and return desired velocity commands. The topic /debug_img1 streams a depth image with an overlaid velocity vector arrow which indicates the model's output velocity command.
The training dataset is available as data.zip (2.5GB, 3.4GB unzipped) from the Datashare. Make some necesary directories and unzip this data (this may take some time):
mkdir -p training/datasets/data training/logs
unzip <path/to/data.zip> -d training/datasets/data
This dataset contains 580 trajectories in various sphere environments. Each numerically-named folder within data contains an expert trajectory of images and telemetry data. <timestamp>.png files are depth images and data.csv stores velocity commands, orientations, and other telemetry information.
We provide a script train.py that trains a model on a given dataset, with arguments parsed from a config file. We provide training/config/train.txt with some default hyperparameters. Note that we have only confirmed functionality with a GPU. To train:
python training/train.py --config training/config/train.txt
You can monitor training and validation statistics with Tensorboard:
tensorboard --logdir training/logs
To create your own dataset, launch the simulation in state mode (after making any desired edits to the chosen environment, camera parameters, or environment switching behavior in the flightmare config file as described in the Test section) to run our simple, privileged expert policy. Note that this included look-ahead expert policy has a limited horizon and is not infallible, occasionally crashing into objects; however, it was used successfully for this paper.
bash launch_evaluation.bash 10 state
The saved depth images and telemetry get automatically stored in envtest/ros/train_set in a format readily usable for training. Move relevant trajectory folders to your new dataset directory. If you have previously cleared the train_set directory, you can do mv envtest/ros/train_set/* training/datasets/new_dataset/. Then, simple edit your config file dataset = new_dataset and run the training command as in the previous section.
We provide a simple ROS1 package for running the trained models on an incoming depth camera stream. This package, called depthfly, can be easily modified for your use-case. On a 12-core, 16GB RAM, cpu-only machine (similar to that used for hardware experiments) the most complex model ViT+LSTM should take around 25ms for a single inference.
You should modify the DEPTHFLY_PATH, self.desired_velocity, self.model_type, and self.model_path in the ROS node python script depthfly/scripts/run.py. Additionally, you need to modify the ROS topic names in the subscribers and publishers as appropriate.
Run a Realsense D435 camera and the model inference node with:
roslaunch depthfly depthfly.launch
We include a /trigger signal that, when continuously published to, will route the predicted velocity commands to a given topic /robot/cmd_vel. We do this with the following terminal command typically sent form a basestation computer. If you Ctl+C this command, the rosnode will send velocity 0.0 commands to stop in place.
rostopic pub -r 50 /trigger std_msgs/Empty "{}"
Some details:
- Keep in mind the model is trained to continuously fly at the desired velocity and would require manual pilot takeover to stop.
- Raw outputs of the node are published on
/outputtopics. - Since it is assumed we are dodging vertical obstacles, we ignore z-velocity model commands and instead institute a p-gain controller to reach a desired height of 1.0 (line 159) (if you decide to accept z-commands from the model then ensure that you read robot odometry and maintain a minimum and maximum height).
- We use a ramp-up in the
run()function to smoothly accelerate the drone to the desired velocity over 2 seconds. - Velocity commands are published with respect to x-direction forward, y-direction left, and z-direction up.
Please fly safely!
@article{bhattacharya2024vision,
title={Vision Transformers for End-to-End Vision-Based Quadrotor Obstacle Avoidance},
author={Bhattacharya, Anish and Rao, Nishanth and Parikh, Dhruv and Kunapuli, Pratik and Wu, Yuwei and Tao, Yuezhan and Matni, Nikolai and Kumar, Vijay},
journal={arXiv preprint arXiv:2405.10391},
year={2024}
}
Simulation launching code and the versions of flightmare and dodgedrone_simulation are from the ICRA 2022 DodgeDrone Competition code.
Error message:
CMake Error: The current CMakeCache.txt directory vitfly/flightmare/flightlib/externals/eigen/CMakeCache.txt is different than the directory <some-other-package>/eigen where CMakeCache.txt was created. This may result in binaries being created in the wrong place. If you are not sure, reedit the CMakeCache.txt.
Possible solution:
cd flightmare/flightlib/externals/eigen
rm -rf CMakeCache.txt CMakeFiles
cd <your-workspace>
catkin clean
catkin build
You can ignore this warning as long as further console prints appear indicating the sending of start navigation command, and the running of the compute_command_vision_based model.
This appears when you are in datagen: 1, rollout: 0 mode, and the scene manager looks for a custom_ prefixed scene to load which is needed for datagen: 0, rollout: 1 mode. You can ignore this warning.