UK-led project aims for AUVs that can learn from mistakes

March 22, 2012 - via The Engineer

A UK-led European project is aiming to build truly intelligent autonomous underwater vehicles (AUVs) that can learn from their mistakes and adapt mid-mission to changing circumstances.

The three-year FP-7 PANDORA project gathers researchers from Heriot-Watt University and King’s College London alongside colleagues from Italy and Greece with industrial input from BP and SubSea7.

Remotely operated vehicles (ROVs) are the workhorses of offshore inspection, repair and maintenance. However, they require up to 6km of cable and the constant presence of a support boat, which is costly.

AUVs have long been seen as a desired piece of technology in offshore environments, however there are some inherent challenges. ‘Once you take the umbilical cable off, you’ve got a problem with talking to the vehicle because electromagnetic waves, radio waves, don’t propagate very far underwater at the sort of frequencies you need to send useful information,’ said Prof David Lane of Heriot-Watt.

Rather than trying to communicate with the robot as ROVs do, the solution is to delegate decision making to the AUV robot itself. ‘It’s quite a tough thing to do because you have to sense where you are relative to the valve, you have to control the forces that the manipulator exerts on the valve and position everything quite accurately — you don’t have a guy there doing that,’ Lane said.

Such AUVs have been investigated for around 10 years now, with some models appearing commercially, although the technology is far from mature. ‘They do okay provided they know the task and the environment very well, but what tends to happen is something isn’t quite right, you’re not in the position you thought, the valve isn’t the way it was supposed to be, then the robot tries to do its mission and fails,’ said Lane. ‘At the moment the systems can’t really deal with that, they cry for help or abort, which isn’t much use to anybody.’

The aim of PANDORA is to build systems that can recognise when they’ve failed. ‘So they’re trying to carry out a task but something has gone wrong — then they can learn from that and figure out how to change what they’re doing in order to succeed,’ Lane said.

The goal at the end of the three years is to produce a prototype that can operate for up to 24 hours, dependent on power constraints and re-charging.

External link:

Related Organizations:
Author:Andrew Czyzewski