Intelligent Workspace Acquisition, Comprehension and Exploitation for Mobile Autonomy in Infrastructure Denied Environments

Vehicles will only get smarter. There will always be a desire for more machine intelligence and autonomy. Our needs and expectations are ever increasing. As a result, we continue to pack more sensors and more computation into the robots that carry, transport, labour for and defend us.

Here we interpret autonomy as a robot's ability to sense, understand and ultimately act of its own accord in its operating environment. This proposal is about giving autonomous vehicles the ability to navigate in difficult conditions over long periods of time. Conditions become "difficult" when GPS is denied or only intermittently available, when little, if anything, is known about the environment, when communications are sporadic and unreliable or when operating conditions like lighting change unpredictably. And yet, somewhat perversely, it is often in just these conditions that our need to navigate is greatest: consider, for example, the surveying of buildings in a stricken nuclear facility such as Fukushima, or the autonomous driving of cars at night in cities where GPS coverage is poor. Intelligent navigation lies at the heart of much of mobile robotics research. It finds application in remote inspection, autonomous urban driving, defence, logistics, security and space robotics.

We shall consider how machines can acquire and manage the information they need to operate persistently in workspaces of our choosing. The goal is to demonstrate that performance improves through use and over time - something that comes naturally to humans and is immensely valuable in a machine. This goal poses questions about how the computers that control robots should represent their environment in a plastic fashion - one which can be stretched and pulled into different shapes over time. We also need to consider how to enable machines to decide how to act to improve their understanding of the world - alone and in concert with other vehicles, each with different sensors and capabilities. How can vehicle sensors be calibrated transparently and continuously? How can motion be planned to maximise both the coverage of inspections and the accuracy of workspace assessments? How can successful operation be guaranteed in the presence of unreliable or short-range communications?

This interweaving of the state of the art in navigation, planning and communications management is unusual and will allow us to ask and provide answers to challenging robotics science questions which, when exploited, will have a dramatic impact on the robots that will become indispensable in the future.

The technical specifications of processors, sensors, networks and storage devices continue to improve at a staggering rate. Driven by this trend, in recent years autonomous systems research has made enormous strides towards applications
of significant value to the public domain. Robots are in the process of revolutionising the running of ports, mines, hospitals, factories and construction sites. Agents which can operate indefinitely will provide another step-change in the impact intelligent systems have on society. It will come in the shape of smarter transport and plant maintenance, lower cost logistics, more planetary science and increased security.

This proposal starts with state-of-the-art competencies in navigation, mapping, multi-vehicle coordination and data-flow management. It asks how these can be used in concert to address our thirst for labour-saving, efficiency-increasing, time-saving, security-giving robots and autonomous vehicles. It sets forth a research program built around the persistent acquisition, comprehension and exploitation of workspace models over indefinite periods, large scales and multiple agents.

To demonstrate impact we have several specific scenarios in mind which encompass a spectrum of autonomy:

1) A Mars rover sample-and-return mission which uses visual navigation to estimate its trajectory and, importantly, allows autonomous path retracing. In many ways planetary science epitomises impact in robotics - there is simply no alternative other than using robots.

2) A consumer, semi-autonomous road vehicle which, for navigation, exploits environmental knowledge acquired from a survey vehicle. In the face of varying lighting and weather conditions, this will be achieved using only a single, low-cost sensor. This approach is in stark contrast to the much vaunted google cars approach, which is always dependent on expensive 3D laser sensors.

3) A heterogeneous fleet of ground vehicles deployed in a reconnaissance role. They must not only navigate in changing conditions, but also plan to maximise the accuracy of their workspace characterisations (e.g. object detection performance) as well as maintaining communications connectivity.

4) Inspection of a nuclear plant or subsea well head with a tele-operated robot. We envision assistive autonomy in which a human operator remains in control of the vehicle but receives guidance on where the camera should be pointed next to ensure a complete and thorough inspection.

We stress that our view of impact is not limited to these four scenarios - they are simply illustrative. What we propose applies to any robotics scenario in which time and time again there is value in knowing "what is where" but one cannot depend on external infrastructure. We do not insist on prior knowledge - we can use it if available but, if not, we will accumulate what is needed autonomously. We do not insist on invariant workspaces - we set out to learn scene dynamics. We do not insist on good communications or superb object detectors - we plan to exploit the platforms' mobility to stay in touch while getting the best views of the world we can.

The mixing of planning, navigation, model-learning and communications management into a single proposal is an unusual, exciting and challenging prospect. It promises impact in the transport, logistics, defence, space and nuclear domains. There will never be fewer robots and what we propose here lies at the core of robotics science and its application to real world tasks.