Decentralized Collaborative Localization with Deep GPS Coupling for UAVs
Siddharth Tanwar and Grace Xingxin Gao, University of Illinois at Urbana-Champaign
There has been an increasing number of applications of UAVs in urban environments for tasks such as delivery, transport, photography, and search and rescue. The United States Federal Aviation Administration (FAA) recorded over 626,000 registered unmanned aircraft systems as of December 2016 and forecasted an annual sale of 5.2 million commercial and hobbyist UAVs in the United States by 2020 . GPS-based UAV navigation in urban environments is affected by error sources such as multipath and signal blockage. On the other hand, with the large number of UAVs available in proximity in the near future, the UAVs will be able to cooperate to achieve better positioning for all agents. However, there are challenges regarding how the UAVs should cooperate with constraints, such as communication load, sensing errors, scalability, etc.
We propose a decentralised collaborative localization framework for UAVs with minimal information exchange and deep coupling of agents' GPS measurements with range and bearing sensor modalities. The framework is applicable to sparsely communicating networks, and information exchange is limited to only those agents which obtain relative measurements. Furthermore, agents carry out updates asynchronously and do not require to store measurement information between subsequent updates. We build our work upon a decentralized extended Kalman filter-based collaborative localization framework . The deep GPS coupling among UAVs enables baseline estimation as well as having virtually visible satellites relayed from the neighbouring UAVs.
In our algorithm, each UAV iteratively updates the belief of its own position and cross-correlation information with the estimates it obtains from its teammates. When two UAVs communicate, they exchange GPS pseudoranges as well as inter-agent range and bearing information along with their beliefs and cross-correlations. Each UAV classifies the GPS pseudoranges into two categories: common satellites visible to both agents, and the satellites visible to only the transmitting agent. Using the satellites in former category, we approximate a single-difference range by estimating baseline between agents. We further augment the baseline estimation by using bearing information obtained from onboard cameras. We estimate pseudoranges to the latter category of satellites from the receiving agent using position estimates of the agents and bearing information. We update these pseudorange estimates by a correction term obtained through linearization of the ranging equation every time we perform a prediction of the belief. These pseudorange estimates, then, augment the set of visible satellites to the receiving agent and are subsequently used in the navigational solution. We incorporate this deep coupling of GPS within the decentralized Kalman filter framework which operates iteratively as follows:
1. Prediction Step – Using a constant velocity model and onboard IMU odometry measurements, each UAV separately performs a prediction step updating the belief and cross-correlation information as well. We obtain correction term and update the pseudoranges for the second category of satellites at this step.
2. Private Update Step - Each UAV computes pseudorange measurements to the satellites visible to itself. It, then, uses this information along with the the pseudorange estimates from the previous step to update its belief and cross correlation information.
3. Relative Update Step – At the instances when two neighbouring UAVs communicate, they perform a relative update step. In this step, we obtain the ranging estimates from the first category satellites (using the method discussed above) and Ultra-wideband (UWB) ranging sensors. Further, we use on-board cameras to estimate bearing information between agents. Subsequently, each communicating agent updates its belief and cross correlation information.
We implement our algorithm on a UAV swarm of quadrotors designed and built by our research group. We setup the swarm with agents flying inside as well as outside a building. While the agents inside the building have a limited view of the sky, the agents outside have a clearer view and hence see more satellites. We have experimentally validated our proposed decentralised collaborative localization algorithm. We have demonstrated the improved performance in terms of positioning accuracy and availability.
 United States Department of Transportation. Federal Aviation Administration. (2017). Retrieved from: https://www.faa.gov/data_research/aviation/aerospace_forecasts/
 Luft, L., Schubert, T., Roumeliotis, S. I., & Burgard, W. (2016). Recursive Decentralized Collaborative Localization for Sparsely Communicating Robots. In Robotics: Science and Systems.