Artificial Intelligence (AI) in robotics: machine learning

My area focuses on the specific applications of robotics to extreme and challenging environments. These include robots that handle nuclear waste, climb tall towers in the middle of the ocean or survive thousands of meters underwater so not vacuuming the floor or serving you a coffee.

AI is cross cutting

AI forms a part of this programme, not because of the current hype around the technology, but simply that some of the latest developments in machine learning are so well suited to robotic challenges in unstructured environments.

Let’s have a look at examples of these latest techniques and the problems they are solving.

Perception: where am I and what am I looking at?

One of the fields of computer vision that has in recent years been disrupted by AI is image classification.

The main advance in recent years is deep learning and particularly convolutional neural networks that are trained to recognise objects in images.

Learning techniques and algorithms for robots

Robot working out sums on a chalkboard

Like many current machine learning techniques, the algorithms were actually first proposed and used on a smaller scale decades ago. The reason for the recent surge in their performance and use is threefold:

So what? In robotics, this is an important capability which can reduce direct or remote human involvement in hazardous environments. A robot capable of identifying objects in images, from an onboard camera, in real time, can perceive more about its surroundings.

underwater vehicle (ROV) being lifted out of the sea
A remotely operated underwater vehicle (ROV)

These algorithms may be trained to recognise specific defects in structures being inspected, potentially using the same transfer learning technique as in some of the recent examples of neural networks for diagnosis from medical images.

Planning and Control: how can I complete my task?

Reinforcement learning is one of the most exciting developments in machine learning in recent years. The most commonly seen examples are those of AI playing various retro computer games.

Three robots working at monitors wearing earphones

These examples may appear to be just for fun however, what they demonstrate is that very general tasks can be solved using only using what the machine can see.

The game’s score acts as the machine’s reward. It means that the algorithms being developed are highly general so they can be retrained for many different tasks and environments.

Child's hand with multi coloured building bricks

In much the same way as a baby will experiment with the world and learn over time what works and what does not, eventually complex behaviours can be learnt in order to achieve a simple goal.

Real life imitating games

So what? If reinforcement learning can be applied to games, it can be applied to simulations of many real-life environments.

aircraft simulator cockpit no pilots, lots of dials and view of runway

If a simulation of a robot and its environment is close enough to real life, then this technique can explore and optimise different solutions to tasks. In some cases, simulations can be sped up so years of learning can be compressed into just hours.

Training a robot workforce

For many robots that work in extreme environments, explicitly programmed instructions are either not feasible, or so time-consuming for a human that the robot becomes an unproductive burden.

Reinforcement learning means that robots can be given simple goals to achieve, and they will use their learnt techniques and their internal knowledge of the world to plan how to achieve a task.

At Innovate UK we are funding projects that use machine learning algorithms across all sectors, and into all of the Industrial Strategy Grand Challenges, from medicine to the digital economy to space, transportation and agriculture.

Robot holding a glowing virtual ball with the words 'AI'

AI points of view

Artificial Intelligence (AI). Two words, which together conjure up an extraordinarily wide range of meanings, and with them, opinions and emotions.

Ask one person, and their view is that AI is an existential threat to humanity; potentially taking all our jobs or even turning against us like the sci-fi like visions of killer robots.

Ask another, and you might be met with an eye roll, that the entire subject can be dismissed as simply algorithms, and anything more is just hype.

Both extreme viewpoints have their truths and their fallacies. In reality, the machine learning techniques that make the modern AI magic tricks happen are progressing at an exciting rate, transforming some industries, but simultaneously the various AI apocalypse scenarios are a comforting distance away.

Prof. Andrew Ng, a prominent AI and machine learning academic has said

Worrying about sentient AI is like worrying about overpopulation on Mars.

To find out more

If you are in the UK and are working in this field to solve robotics problems with the latest machine learning technologies, get in touch to find out how your academic or business R&D can be supported by UKRI.


You can follow Mark on Twitter

You can follow Innovate UK on: