Robotic GNOME Laboratory

Long-Term Autonomy and Persistent Navigation in Spatiotemporally Dynamic Environments

Ocean imagesNavigation diagram

The goals of this project are to overcome the theoretical and technical challenges of developing a general prediction, control and planning framework for autonomous navigation and sampling of dynamic ocean features. PI Ryan N. Smith will make contributions to the areas of navigation, deliberation, prediction, and targeted sampling to extend autonomy in marine robotics. In pursuit of this goal, the objectives of this proposal are to develop the broad principles that enable autonomous aquatic vehicles to: 

  1. Improve navigation and localization for operation in dynamically-evolving environments,
  2. demonstrate the utility of prior information in planning and deliberation,
  3. investigate human in-the-loop control strategies and the associated data analysis and transmission necessary, and
  4. experimentally validate research outcomes through simulation, laboratory tests, and field trials.

OBJECTIVES

To achieve the long-term goals outlined for this project, the PI will make contributions to five important research objectives. These contributions align to tackle two important problems 1) the coordination and control of multiple heterogeneous platforms, and 2) the localization of the team for targeting sampling efforts.

Research Objective 1: Novel prediction and modeling strategies will extend the notion of locations in the ocean existing in geographic space to instead be drawn from, or exist in an environmental space that defines the fluid dynamics intrinsically. Increasing use of AUVs for scientific exploration will require weaning away from traditional straight line transects. To observe fine-scale biogoechemical processes, variable-speed vehicles will need to track advected patches of water and provide high-resolution environmental context for these dynamic features (e.g., fronts, plumes of different kinds, anoxic zones). The technical challenge here is to be able to describe and flexibly move waypoints based on the predicted displacement of a moving water mass. The critical change is the definition of waypoints; from a world reference model to one relative to the moving water mass or feature. Further, these waypoints must also depend on the position and speed estimates of the vehicle relative to the dynamic feature. This cannot be known a priori, and must be computed in situ [1,2].

Research Objective 2: Gathering data of scientific interest is a difficult task, and sensors must be in the right place at the right time for targeted ocean sampling. Deciding where to sample within a large stochastic field is a challenge. Basing decisions on uncertain model outputs with the goal to refine the models adds to the challenge. We address this by using a combination of Entropy driven methods and on-board adaptive intelligence using Artificial Intelligence (AI) Planning/Execution with Machine Learning for environmental state estimation.

Research Objective 3: Reactive control strategies on-board the vehicles will enable adaptation to a spatiotemporally evolving environment. Contrary to terrestrial applications, aquatic platforms generally experience poor navigational reliability, as GPS, VIO and SLAM are significantly limited. Additionally, ocean currents are typically the same order of magnitude of the vehicle velocity, causing the uncertainty in dead-reckoning solutions to increase without bound. Here, we investigate how to conduct route planning and plan execution in the context of predicted uncertainty to improve mission execution in dynamic environments.

Research Objective 4: Throughout a sampling mission, what decisions are best made on-board, and which should utilize a human in the decision-making loop? While full-autonomy is a continuing goal, some aspects of automated inference will remain onshore to augment human-in-the-loop control methods. Mixed-initiative methods of control [3,4] will therefore need to be explored in such contexts and with persistent uncertainty in communication-limited aquatic environments.

Research Objective 5: Ocean sampling is an exemplary candidate for the application of developments in machine and/or deep learning, specifically in the context where big-data is analyzed but represents only a sparse sampling of the environment. Recently, with the advent of robust robotic platforms, the deluge of information from multiple sensors has created an opportunity for the rise of methods derived from Information Theory [5]. This is particularly pronounced in marine robotics where a fundamental problem has been in dealing with large spatial scales and where to sample the complex ocean is driven by statistical methods derived predominantly from the agricultural and environmental sciences. Gaussian Process [6,7] and Deep Learning [8] are approaches that will play a dominant role in ocean exploration and marine autonomy.

RESULTS

Task 1: Through multiple deployments in Big Fisherman’s Cove on Santa Catalina Island, CA we have begun to examine the joint relationship between satellite imagery of ocean features, collected science data, and bathymetric maps of given regions. In Figure 1, we show a satellite image of the deployment region on the left, with the interpolated area delineated by the black polygon. On the right in Figure 1, we plot an interpolated map of the sum of the normalized values (0-1) of the surface temperature, surface dissolved oxygen, and water depth. We see a correlation between the collected and interpolated data and the satellite image that delineates the reef, sand, and other bottom type. This naieve approach to formulating a joint relationship provides motivation to further examine this idea. Additionally, we have applied an adaptation to the terrain-based navigation approach presented in [9] to a set of science data collected along a path that was not included in the interpolated map with promising results. The idea is that by combining relevant science data into a single, 2-D scalar field, we can create a underlying map that would have enough variability to be utilized in a similar fashion to a bathymetry map for terrain-based navigation. We are examining the utility of such a map with and without depth as an included parameter. The significance of this result is that we may be able to localize relative to collected science data, and follow water patches dynamically rather than rely on a global coordinate system for navigation.

Satellite image of Big Fisherman’s Cove and weather data

Figure 1: Satellite image of Big Fisherman’s Cove on Santa Catalina Island, CA (left). Normalized sum of gathered data for temperature, dissolved oxygen, and depth plotted over the same area.

Task 2: Through prior collaboration with Drexel University, we have developed a stochastic planning strategy for AUVs that enables adaptation and reaction to a spatiotemporally evolving environment. This strategy was demonstrated in a laboratory testing tank with scaled versions of surface and underwater vehicles [10]. We have successfully integrated the developed algorithms onto the two surface vehicles and the underwater vehicle acquired through the NSF MRI mentioned below. The significance of this result is that we are now able to seamlessly go from laboratory to field trials with developed algorithms.

Task 3: No results to report at this time.

Task 4: We have successfully conducted multiple field deployments in the controlled environment of a small reservoir (Rogers Reservoir) and in the ocean (Santa Catalina Island) with the Heterogeneous Networked Instrument for Aquatic Exploration and Intelligent Sampling. We have recently gaied access to Lake Nighthorse, a 6 km2 reservoir with depths exceeding 40 m that is currently closed to public access, and have completed one deployment. Both of the local reservoirs (Rogers and Nighthorse) are fed from local rivers with pumping systems. We have negotiated control of these pumps to create dynamic plumes within the reservoirs to test and validate developed algorithms and sampling strategies for spatiotemporally dynamic events. An initial trial was conducted in Rogers Reservoir, and the temperature data from a lawnmower survey are presented in Figure 2. Note the dark blue signature of the new water coming into the reservoir. As this water was colder than the existing water, this surface survey did not capture much of the incoming plume. The significance of these results is that we have an operational fleet of vehicles, and are able to generate dynamic, acquatic features on command in the laboratory, and mulitple field scenarios. This creates a pipeline for experimental validation of algorithms and sampling strategies across a range of operational scenarios.

Temperature data chart

Figure 2: Temperature data collected during a lawnmower survey of a plume created by pumping new water into the reservoir.

REFERENCES

[1]  J. Das, F. Py, T. Maughan, M. Messie, T. O’Reilly, J. Ryan, G. S. Sukhatme, and K. Rajan, “Coordinated sampling of dynamic oceanographic features with auvs and drifters,” Intnl. J. of Robotics Research, vol. 31, no. 5, pp. 626\–646, 2012.

[2]  J. Das, F. Py, T. Maughan, T. O’Reilly, M. Messi, J. Ryan, G. S. Sukhatme, and K. Rajan, “Simultaneous tracking and sampling of dynamic oceanographic features with auvs and drifters,” International Journal of Robotics Research, 4 2012.

[3]  J. Bresina, A. K. Jonsson, P. H. Morris, and K. Rajan, “Activity planning for the mars exploration rovers,” in International Conference on Automated Planning and Scheduling (ICAPS), 2005.

[4]  M. Ai-Chang, J. Bresina, L. Charest, A. Chase, J. C.-j. Hsu, A. Jonsson, B. Kanef- sky, P. Morris, K. Rajan, J. Yglesias, B. G. Chafin, W. C. Dias, and P. F. Maldague, “{MAPGEN}: Mixed-initiative planning and scheduling for the mars exploration rover mis- sion,” IEEE Intelligent Systems, vol. 19, no. 1, pp. 8–12, 2004.

[5]  T. M. Cover and J. A. Thomas, Elements of Information Theory. Wiley Online, second ed., 2005.

[6]  P. Rigby, O. Pizarro, and S. B. Williams, “Toward adaptive benthic habitat mapping using gaussian process classification.,” Journal of Field Robotics, pp. 741–758, 2010.

[7]  R. Graham, F. Py, J. Das, D. Lucas, T. Maughan, and K. Rajan, “Exploring space-time tradeoffs in autonomous sampling for marine robotics,” in Intnl. Symp. on Experimental Robotics (ISER), 2012.

[8]  I. Arel, D. C. Rose, and T. P. Karnowski, “Deep Machine Learning - A New Frontier in Artificial Intelligence Research [Research Frontier],” Computational Intelligence Magazine, IEEE, vol. 5, pp. 13–18, Nov. 2010.

[9]  A. Stuntz, J. S. Kelly, and R. N. Smith, “Enabling Persistent Autonomy for Underwater Gliders with Ocean Model Predictions and Terrain-Based Navigation,” Front. Robot. AI, vol. 3, p. 23, Apr. 2016.

[10] D. Kularatne, M. A. Hsieh, and R. N. Smith, “Zig-Zag Wanderer: Towards Adaptive Tracking of Time-Varying Coherent Structures in the Ocean,” in IEEE International Conference on Robotics and Automation, 2015.

[11] D. Heermance, D. Kularatne, J. D. H. Sosa, M. A. Hsieh, and R. N. Smith, “Design and Validation of a Micro-AUV for 3-D Sampling of Coherent Ocean Features In-Situ,” in MTS/IEEE Oceans, 2015.

PUBLICATIONS FOR THIS PROJECT