THEORETICAL MODELS
3.2 BACKGROUND
We constructed a set of simulations that explored the feasibility and benefits of adaptive planning for Mars exploration. These simulations combined high- and low-level planning route UAVs among Mars localities of varying scientific value, subject to fuel and time con-straints. The high-level planner performed site selection according to scientific objectives in tandem with a low-level kino-dynamic path planner that transmitted waypoints for UAV routing. The full experiment is described in Appendix B, and summarized below. These simulations showed that the adaptive planning approach identified solutions that optimized the value of information within the constraints of the simulation.
Briefly, this experiment assumes a set of location of unknown information value. The adaptive mission planner calculates an exploration strategy that maximizes utility within fuel constraints. The plan is then transferred to the kino-dynamic path planner, which cre-ates waypoints to guide UAVs to the locations in the most efficient manner. Way- points are designed based on a one-second time scale, so as to discretize each path unit on a relatively fine scale. We interfaced the adaptive mission planner with the Mars Plane simulator. The adaptive mission planner constructs a plan and calculates the fuel usage. Excessive fuel
Figure 2-4: The graph shows the results results of plan execution for three strategies. The rate of change is along the x-axis and the average total utility gained is on the y-axis. The blue line is the static finite-horizon plan, the green line is the adaptable finite-horizon plan and the red line is the greedy plan (adaptable one-step receding-horizon plan). The static planner outperforms the greedy planner under stable conditions and the greedy planner outperforms the static planner under unstable conditions. The adaptive planning strategy performs well in all situations. It performs at the same level as the static planner under stable conditions, outperforms either planner under moderately changing conditions, and matches the performance of the greedy planner under very unstable conditions. (See Ap-pendix B for more details)
usage may lead the planner to construct and alternative plan. Figure 2-5 shows the oper-ator interface for the simulated system. The simulated environment includes geographic features, such as mountain ranges, that the UAVs should avoid, as well as craters, outcrops and alluvial fans that have potential information value.
The sites are color coded in Fig. 2-5 along with the UAV starting location (a white square), a set of utilities for each science site (yellow numbers) and the travel distance implicit in the map area. As the planner receives the actual travel time to the current site and the updated utilities, it creates a new plan (Fig. 2-4 e and f), which optimizes utility and fuel use according to the new information and constraints. In this situation, the adaptive planner can accrue greater utility while operating within the fuel constraints relative to that accrued by a static plan across a range of environments that were changing at different rates (see Appendix B for further details.)
2.4 3-level Hierarchy
The next set of challenges in framing a planetary survey in terms of an info-MDP con-cern how to cover a large geographic area and integrate multiple sensor types in the active sensing task. We designate a 3-level spatial hierarchy that allows us to address the unique spatial and sensor integration challenges that arise in sample spaces of different scales. The 3-level hierarchy shown in Figure 5-3 includes a synoptic level, which represents a wide area (103 km or more), a mesoscale level, representing an intermediate-sized area and a microscale level, covering a small, local area (less than one km). This hierarchy allows us to implement an integrated system of info-MDP algorithms to adaptively survey the sample space according to basic information objectives.
The instruments, vehicles, objective and operations at each level are different, but may overlap. A satellite provides coverage and support at the synoptic level, while UAVs operate at the meso- and microscale levels (Fig. 5-3). The satellite coarsely surveys a large swath of the terrain and then relays information to a group of UAVs. The UAVs deploy to the areas that appear to have high value of information. Each level requires a different optimization strategy due to the different sensor functionalities and the abstraction of the environmental
(a) (b)
(c) (d)
(e) (f)
Figure 2-5: Mars Airplane Adaptive Mission Planner: The ARES aircraft could explore a diverse terrain in the equatorial Valles Marineris region shown in (a) at a vantage point 1.5 km above the surface. Our high-level adaptive mission planner performs site selection based on an evolving science value estimate. A lower level kino-dynamic path planner executes the resulting plan. The map shows the aircraft entry point and 3 categories of sites to explore. (b) The numbers indicate the science value estimate (utility) of the 3 site categories. (c) The Adaptive Mission Planner uses these utilities and fuel constraints to construct a plan. The kino-dynamic path planner executes the plan. When the aircraft visits and observes a site, it updates the utility estimate for that particular site category.
Panel (d) reflects the effect of observing a blue site. The adaptive Mission Planner uses the new utility values to formulate a new plan. (e) The Adaptive Mission Planner also updates its time horizon based on the actual distance traveled. A significant deviation from the estimated distance results in a new plan. In this example, the Mars airplane took longer than expected to reach a planned site and a new, shorter plan was constructed. (f) Our system includes a low level path planner that can simultaneously avoid obstacles (weather,
Figure 2-6: The Autonomous Mars Exploration Network: We extend the notion of au-tonomously exploring Mars to a multi-level framework incorporating as satellite surveyor, which provides synoptic information to a team of UAVs. These UAVs explore the surface of Mars for features of scientific interest. Once located, the UAV can map out a set of traverses that can be executed by a ground rover to the site.
Figure 2-7: The 3-level hierarchy shown above includes a synoptic level (covering a wide area), a mesoscale level (covering a medium sized area) and a microscale (covering a small local area). The synoptic sensing (satellite) algorithm concentrates on dwell time and the value of the collected information. The mesoscale system is optimized with respect to vehicle travel time. Travel and dwell time are simultaneously optimized at the microscale during free-form AUV exploration.
model. A satellite can quickly refocus its survey within its sensor footprint. The sensing algorithm therefore concentrates on dwell time and the value of the collected information.
The dwell time refers to how long the satellite looks at a particular location. The satellite is therefore optimized to focus on areas of high interest and high uncertainty.
At the mesoscale level, the UAVs transit directly to the locations of interest. The infor-mation gathered during transit is considered negligible, and is ignored in the formulation, thus simplifying the model of reward. The mesoscale system is optimized with respect to travel time and fuel constraints. The satellite provides a set of target areas that are divided up among UAVs according to an optimized vehicle routing plan. The microscale plan-ner handles the local sensing tasks. At the microscale level, the UAVs perform free-form exploration, collecting data as they move. Travel and dwell time are thus simultaneously optimized at the microscale. Next we examine and formulate the info-MPD for each levels of the hierarchy.
Figure 2-8: Satellite-UAV collaboration can be broken into different levels. At the mission level, we route the UAVs to the mapping task locations. In the local level, each location is its own information gathering task with the goal of estimating key environment variables in the area. Both problems can be solved with an uncertainty guided planning algorithm.