Sustained Autonomy through Coupled Plan-based Control and World Modelling with Uncertainty

Sustained autonomous behaviour requires a system that is robust to uncertainty, both at the low level of interactions between actuators and sensors and its environment, but also at the intermediate level of sensory perception and interpretation, action dispatch and execution monitoring and, at the highest level of planning, action selection, plan modification and world modelling. In this project we bring together a team of experts with complementary and linked skills and experience, from robotics and sensor data processing, from planning and from reasoning under uncertainty. Our goal is to combine these areas in order to build and demonstrate a robust approach to sustained autonomy, coupling plan-based control to the construction of world models under constraints on resources and under uncertainty.

We plan to demonstrate the approaches in an underwater environment, using Autonomous Underwater Vehicles (AUVs), performing inspection and investigation missions. These missions share many features with space exploration, the use of autonomous Unmanned Aerial Vehicles (UAVs) for track-and-target missions and investigation of terrestrial hazardous sites, such as nuclear waste storage sites. In all of these case, communication between a human supervisor and the autonomous system is often tightly constrained. This is particularly true of deep space missions (for example, Mars missions face transmission delays of about 15 minutes, but windows might consist of a just two 30 minute slots in 24 hours). However, in aerial observation missions involving multiple assets the need for rapid responses in the control of fast moving vehicles reacting to agile targets also leads to bandwidth constraints for a single human controller attempting to manage and coordinate the mission. Hazardous sites, particularly those subject to radiation emissions, often contain communication black-spots where vehicles must operate without human intervention over extended periods. The underwater setting also imposes limits on communication due to the physical difficulties in transmitting control signals over significant distances.

Many of these missions involve multiple assets, often mounting different capabilities. Space missions might combine orbital observing assets, ground-based landers or rovers, possibly aerial vehicles (in some settings) and even astronauts, each offering different subsets of capabilities. Aerial observation might combine slower but more agile vehicles with others that are fast but less manoeuverable, while mounted imaging systems might exploit different wavelengths (visible, infrared, radar) and vehicles might offer other capabilities. We intend to explore the use of multiple assets, including the coordination of AUVs mounting different sensors and actuators.

Uncertainty offers different challenges according to environment. Many space environments have relatively predictable dynamics (although Martian winds are one example of a highly dynamic and uncertain factor), but aerial observation missions operate in highly dynamic and unpredictable environments: both atmospheric conditions and target behaviours can be a source of dynamic uncertainty. The underwater environment is also highly dynamic: phenomena such as currents will act as useful proxies for similar dynamic sources of uncertainty in other execution environments. Other sources of uncertainty arise from the inherent limitations of sensors and actuators and our ability to process and interpret the data that can be recovered from these devices. One of the biggest challenges in achieving robust autonomy is in recognising that uncertainty about the state of the world and the state of execution of a plan is inevitable, but the form of that uncertainty is itself an unknown.

By combining techniques in modelling and reasoning about uncertainty, plan modification and sensor data perception and interpretation, we propose to build a robust approach to autonomous systems control.

Autonomous functioning of complex systems is becoming an increasingly widespread problem of interest, from the management of specialised missions (space, tracking, inspection, military, rescue and so on) to management of infrastructure (energy, transport and water), the objective is to improve efficiency, responsiveness and reliability without increasing costs. The consequence is a huge increase in interest in technologies to support autonomous decision making and intelligent control. The industrial sponsors for this Call for Proposals represent an important subset of companies in the forefront of the drive for this technology. It is clear that the competitiveness of the UK will depend on leading in the development of these technologies for intelligent automation, robotics and autonomous control.

Our first pathway to impact is therefore clear: we will work closely with the sponsors of this call to develop solutions to intelligent control under uncertainty that are close to deployment. The structure of our workplan is designed to facilitate this, with a 6 month lead in during which we will ensure the closest fit between our reseach directions and the partners' goals and a 6 month conclusion during which we will aim to transfer the ideas from the project into the partner organisations. This will be an excellent and organised opportunity for exploitation and will provide a route to embedding the technologies within the labs of the sponsors.

Beyond this, all partners have links with a variety of companies and organisations that will be interested in the further development of these technologies. Fox and Long have connections with the power industry through project partners in Electrical Engineering at Strathclyde, including Scottish Power, where autonomous smart grid operations are seen as an essential future technology. Fox is adjunct researcher at Monterey Bay Aquarium Research Institute, offering opportunities for pushing this technology into the oceanographic community. Dearden has close associations with the Autosub group at Southampton, offering similar opportunities and Lane is a member of the Marine Alliance for Science and Technology in Scotland (MASTS) as well as enjoying a wide range of links into the marine technologies community, so there will be a wide range of possible avenues open to partners for passing this technology into relevant areas of application in underwater operations. Long has good connections with Steve Chien, Chief Scientist at JPL, working on autonomy for space, including AEGIS and EO-1. The work being proposed in this project will be of interest to Chien and his group and Long has regular contact with Chien, including through formal meetings such as the International Workshop on Planning and Scheduling in Space and the IJCAI Workshop on AI in Space series by which the work can be disseminated.

The impact we expect that this project will have is in defining and demonstrating new ways to manage uncertainty thoughout the architecture of an autonomous system, from the sensor data interpretation, through world modelling, to the planning and plan modification in response to the effects of uncertainty. We also envisage that the work on V&V for autonomous behaviour in an uncertain environment could have broad impact, particularly amongst existing users of technologies for autonomy, such as JPL and NASA. Current approaches to V&V depend on extensive testing and software V&V techniques such as model checking. New ideas in this area are urgently needed and any successes would have significant impact.