Augmented capabilities of UAV drone units

One of the tenets of the Drone video sensing platform we have described in the previous articles is that the platform makes no assumptions about the capabilities of the UAV units or their sensors other a camera capable of First-person view aka FPV drones. Let us review what these additional capabilities mean in terms of algorithms. We take the case when UAV swarm are capable of autonomous flight and inter-drone communication. This provides an opportunity for additional layer of computing over centralized and decentralized task assignments via waypoint tracking and self-organizing maps, respectively. In this layer, each drone follows a systematic process to achieve its objectives and collaborate within the swarm.

 We can safely delegate the communication to a mesh network and collaboration with consensus protocols such as Paxos or Raft and instead focus on the computing plane shortly. The Mesh network is decentralized and relaxed enough for the drones to join or leave the network as they move in and out of range; the network itself is self-healing with redundancy for reliable communication over multiple paths and scalable without degradation in performance. If some form of storage were available, cachepoints arranged in a ring can play a part in consistent hashing over a distributed hash table with data versioning to enforce consistency. They also work well with consensus protocol especially those that implement state machine replication that combines transaction logging for consensus with write-ahead logging for data recovery. When these state machines are replicated, they become fully Byzantine tolerant.

In the computing plane, these capabilities provide an opportunity to improve performance by providing the necessary information for agents to select an action. With the payload in exchanges to not be limited to any one type of sensor and getting reinforcement learning from adjoining drones, an adaptive framework can be designed to significantly boost the efficiency of critical tasks such as reconnaissance and collision avoidance. If the collective thinking of the UAV swarm needs to be overridden, a single drone can be targeted to move in the desired direction with a velocity to automatically move the group to follow. Since this computing builds upon decentralized framework without impacting the earlier self-organizing maps, it forms a clean layer of computing above the existing to enhance the dynamic behavior of the swarm and its error corrections.

The systematic processes and reinforcement learning referred to earlier for each drone must include those for inter-drone distance calculations such as Euclidean distance, a minimum spanning tree to care for all the drones in the swarm by creating a graph of drone distances and policy optimization to learn optimal behavior to achieve a given goal by adjusting search patterns from latest data and updating policies while maintaining learning stability. With a frequent iteration of updating positions of each drone, recalculating inter-drone distances between pairs, and constructing the MST by merging the smallest edges between the fragments until all fragments are connected, the UAV swarm can respond faster and better to meet critical operational requirements such as collision avoidance. Additionally, if velocity were to be captured and shared along with the position information, the model can accommodate dynamic changes and better predict agent motion and path planning. Setting a suitable threshold in the policy optimization allows us to counter instability and abrupt policy changes. A classifier that uses the policy optimization loss can ensure coherence for the entire swarm even under varying conditions which significantly boosts reliability of decision making. Inter-agent co-operation is in this way, a desirable layer of computing that the UAV swarm platform can bring to the table when augmented capabilities are available on the drones.