This is experimental firmware, use at your own risk!

There is still a significant amount of research required to validate the behavior of neuro-flight controllers. If the neural network is not trained properly, instabilities and unpredictable behaviors will occur.

Neuroflight is the first open source neuro-flight controller software (firmware) for remotely piloting multi-rotors and fixed wing aircraft. Neuroflight's primary focus is to provide optimal flight performance.

Neuroflight aims to address limitations in PID control used in Betaflight through the use of neural network flight control (neuro-flight control). Neuro-flight control has been actively researched for more than a decade. In contrast to traditional control algorithms, neuro-flight control has the ability to adapt, plan, and learn. To account for dynamic changes Betaflight has introduced gain scheduling to increase the I gain when certain conditions are met, for example low voltages or high throttle (anti-gravity). On the other hand, neuro-flight control learns the true underlying dynamics of the aircraft allowing for optimal control depending on the current aircraft state. For example neuro-flight control has the potential to learn the batteries discharge rates to dynamically adjust control signal outputs accordingly. The goal of this work is to provide the community with a stable platform to innovate and advance development of neuro- flight control design for drones, and to take a step towards making neuro-flight controllers mainstream

• 2018-11-14 Stable flight has been achieved with Neuroflight https://youtu.be/c3aDDPasjjQ

In addition to features provided by Betaflight 3.3.3,

• Neural network based flight control
• Mixing replaced by neural network

Interfaces are defined for the sensors available on hardware. As models become more sophisticated additional sensors will be used (e.g. ESC telemetry, voltage sensor, etc.)

Input (x) is of size 6, where x = [roll error, pitch error, yaw error, delta roll error, delta pitch error, delta yaw error]. Inputs are unbounded and in degrees/s. Output (y) of size N corresponding to motor 1 ... motor N. Each output value is in range [-1, 1].

1. Use GymFC to train and create a neural network in the form of a Tensorflow checkpoint.

2. Place the checkpoint files (four of them: checkpoint, *.data, *.meta, *.index-*) in a directory which can be independently version controlled.

3. Create a file called tf2xla.config.pbtxt in the directory and define the neural network configuration according to https://www.tensorflow.org/xla/tfcompile.

4. Neuroflight was developed using Tensorflow-1.8.0. There appears to be a bug/issue in Tensorflow-1.8.0-rc1 preventing the ABI type from being passed to tfcompile used to compile the neural network. A quick hack to force the correct ABIType is to modify compiler/xla/service/llvm_ir/llvm_util.cc. At the end of the function SetTargetOptions place,

target_options->FloatABIType = llvm::FloatABI::Hard;

Need to investigate whether these bugs have been fixed in newer versions or come up with a better method to handle this. Install Bazel and then build Tensorflow.

5. There appears to be a second bug in which tensorflow/compiler/aot/runtime.cc does not import malloc.h.

1. Set the environment variable TENSORFLOW_DIR to the location where you have installed Tensorflow. Best to export this in your bashrc or you can put this in your local make file, make/local.mk.

2. In make/local.mk define the CHECKPOINT variable for the directory containing your neural network checkpoint.

3. Refer to Betaflight or src/main/target for list of supported FCs.

Any F7 should be fine. Flash memory is sufficient for F4's however it is unknown how the decreased processor speed will affect execution of the neural network.

• Matek F722-STD

Neuroflight is compatible with the Betaflight Configurator however any modifications to the PID controller and mixer will not do anything as they are not used by Neuroflight.

In order to reduce maintenance, avoid merge conflicts and keep as in sync with upstream Betaflight, Neuroflight's architecture will maintain a minimize footprint and isolate its code from Betaflight as much as possible. The following table describes the files modified (M), and added (A) for Neuroflight,

At the time of development Neuroflight was forked from Betaflight 3.3.3. Since then, newer versions of Betaflight have increased in size as more features are packed in. Further testing is needed to identify if the neural network will fit in newer versions of the firmware.