Investigators: Eric Frew
Collaborator: Han-Lim Choi, Korean Advanced Institute for Science and Technology (KAIST)

The purpose of this collaborative effort is to develop and evaluate a hierarchical framework for real-time, path planning for persistent information-gathering tasks in fading communication environments. Novel features of the proposed work are learning the radio environment and background winds; consideration of the energy that can be gained from wind field patterns (energetics); and a persistent, communication-aware information-based formulation of sensing tasks (informatics) that can integrate multiple sensing modalities.

Persistent autonomous information-gathering tasks require computationally efficient algorithms that can plan sensor motion while considering i.) dynamic feasibility; ii.) energy extraction from strong wind fields or currents; iii.) communication limitations due to stochastic fading of radio transmissions iv.) sensor fields of view; and v.) multi-modal sensor capabilities. Cooperative control of multiple unmanned aircraft has the potential to allow persistent information-gathering of endurance and quality that could not be achieved without coordination. However, cooperative control is complicated by stochastic fading communication channels that degrade information-gathering performance. Such limitations favor keeping aircraft close together. On the other hand, wind fields provide an energy source that can be exploited by the planning algorithms to prolong system endurance. However, this exploitation process introduces new trade-offs between map-building, energy harvesting, and information-gathering tasks, often encouraging aircraft to move apart from one another. The dynamic, uncertain nature of both the targets of interest for the information-gathering mission and the wind and radio environments necessitate path planning algorithms that can easily adapt and replan. As a result, sample-based planning methods are preferred since they give feasible results quickly that can be improved given more time and since they build search trees that can be saved for re-use with minor modification compared to optimal methods that require computationally-expensive, exhaustive techniques.

In the context of the overall planning framework, the RECUV team will develop machine learning algorithms for mapping the environment and will develop fast sample-based planning algorithms that exploit the maps during information-gathering tasks. Gaussian processes will be used to derive probabilistic environment models. Information-theoretic concepts will be used to formulate communication-aware information-gathering objectives that can integrate multiple sensor modalities, e.g optical sensors and radar. The parallel implementations of sample-based algorithms will be extended to include the dynamic models of system energetics and informatics and to enable fast replanning once an initial plan has been created.