NXP HoverGames Challenge 2: Help Drones, Help Others During Pandemics with NXP
Checkout our Wiki Here!
Wild animals destroy property, kill/infect livestock, as well as harm agriculture. We will create a drone that can herd wild animals, preventing such harm.
See it here!
In this repo you will find our C++ code for our drone and our ground station as well as a python interface for our ML code developed by Adam.
Our system has three first class citizens:
- The Agent(s) (drone) - a node.
- The Base Station - the central, main, node.
- The Controller(s) - a node.
The scene consists of two players:
- Agents - drones to direct to herd animals out of zone.
- Subjects - animals to keep out of keep out zone.
There can be many agents, and many controllers, however there is only one centralized base station. A controller runs the RL algorthm and subject classifier, and is in ownership and control of the agents in the scene. An agent is owned by one controller. The base station acts as a router/conduit and rule enforcer between the controllers and the agents. The base station is responsible for setting up a network on top of UDP/TCP over WiFi depending on the configuration. This network has a defined set of messages that can pass through it. When a new node is created, it asks the base station for an ID and lets the base station know what type of node it is (Agent or controller). Once the agent has been assigned a node ID it is able to begin its life.
The RL algorithm and the animal classifier/tracker run on the controller node. As the classifier neural network detects subjects (using humans as a stand in for testing) it fills the RL scene with subjects. Additionally a controller can share its subjects with other controllers by sending a message to the base station. Meanwhile the drone is constantly broadcasting its location to the various controllers by way of the base station. The RL algorithm runs over the scene, processing the locations of the subjects and the agents to optimize the commanded position of the agents it owns. The base station then generates a target location for the agent and the agent then takes off and flys to that location. As the agent travels to the location our RL algorithm constantly tracks the subjects and the agent which enables the algorithm to change instructions on the fly. Then as the agent approaches the subjects they are scared away and out of the keep out zone. Once the subject of interest has left the keep out zone the agent lands back in its starting position and awaits new instructions.
Here you can find our RL + Neural Network related code for:
- Detecting animals and
- Directing agents (drones) to herd the detected animals.
Note about my findings here