Onboard Adaptive Control of AUVs
MBARI’s Autonomous Systems Group, led by Kanna Rajan, has been working with scientists, engineers, and Marine Operations personnel to augment autonomous underwater vehicles (AUVs) with machine intelligence.
Until recently, the capability to autonomously detect and sample oceanographic features of interest did not exist. To establish this important capability, MBARI’s Autonomous Systems Group, led by Kanna Rajan, has been working with scientists, engineers, and Marine Operations personnel to augment autonomous underwater vehicles (AUVs) with machine intelligence. MBARI routinely uses AUVs in gathering water-column time-series data as well as bathymetric data for detailed maps of the ocean floor.
Machine intelligence offers advancements in not only scientific research, but also AUV operations. AUVs rely on battery packs with limited endurance and, while submerged, do not typically communicate with a support ship or shore facility. Mission plans have to be scripted a priori using simple tools on shore with the expectation that the vehicle will rigidly execute the plan. In contrast, onboard machine intelligence permits dynamic adjustment of mission plans to enhance the utilization of the AUV’s resources, considering mission objectives, environmental conditions, and resources available.
To enable new modes of scientific research and AUV operations, Rajan’s team is incorporating advanced Artificial Intelligence techniques into an integrated vehicle control system called T-REX, for “teleo-reactive-executive”. T-REX is built around the paradigm of sense-deliberate-act. The AUV uses its instruments to sense the environment, it generates a plan or course of action for the future by deliberating about the best way to accomplish the goals, then it acts on (or executes) this plan. The software integrates a number of science observations to estimate the likelihood that the vehicle is inside a feature of interest. Deliberation of mission goals allows the software to reason how best to carry out the specified scientific goals, taking into account the dynamic ocean environment and the potentially changing state of the AUV itself, such as diminishing battery power. The control system then validates the data and can quickly decide to investigate the phenomenon more closely. It does so by changing its navigation parameters to ensure it can observe the phenomenon at closer range and at a higher resolution. Should there be an instrument failure or a short in the battery that would impact the mission goals, the system generates a contingency plan to work around it.
How the AUV reaches a goal will evolve
as does the plan itself. Technically, T-REX
breaks such goals down into manageable
problems it knows how to solve. It does so
by consulting a model that describes the
interactions of the various subsystems within the vehicle.
This model allows for complex relationships between various
parameters using a network of constraints. It searches likely
outcomes of the plan while permitting the specification
of constraints, such as “If we want to use the Global Positioning
System, we must be at the surface,” which triggers
further planning activities (i.e. a goal is implied to get the
AUV to the surface). Such goal-directed behavior enables
a scientist to tell the AUV what to do without specifying
how to do it. The how is encapsulated in an AUV model
which T-REX consults to “do the right thing”.
A database stores these variables and constraints and is used
to incrementally generate a set of activities that become
a sequence of steps in the plan. If any of the hard and fast constraints are not satisfied, then the planner has to “backtrack” and look for alternative choices. Typically the system rapidly solves upwards of thousands of such constraint equations in less than a second within the control loop of the system.
The long-term vision of the Autonomous Systems Group
at MBARI is to use advanced automated techniques for
sustained ocean exploration for water-column as well as
seafloor measurements. Imagine a future in which very
capable AUVs roam the ocean, and in the process of conducting
routine observations, adapt to opportunistic science
discoveries that alter the mission plans dynamically, all
without human intervention. Like well-trained bird dogs,
fleets of AUVs “sniff” out interesting science and retrieve
relevant samples at a small fraction of the cost of oceanographic