IET Central London Evening Lecture, Savoy Place, 11-May-2016 

Professor Yiannis Demiris, Personal Robotics Laboratory, Imperial College, London


Are robots our enemies? Will they take over the world from humans? 

Not at all! As long as engineering teams like this one at Imperial College prevail.

Their ethos is to contribute to the betterment of the human race: Not by replacing humans, but by assisting us to do things better, particularly where we have disabilities or recuperation challenges. They have learnt not to try to help people who don't want help.


Their robotic applications include such areas as:

• Smart robotic wheelchairs for kids with disabilities. 

• Reinforcing rehabilitation exercises. 

• Assisting people with dressing. 

• Learning physical tasks such as dancing; the robot adapts to what a kid struggles to do - to strengthen those movements. 

• Creative tasks - such as musical collaborations

The key issues for optimal robot assistance are that it must be predictive (not learning from errors, such as falling down!) and be able to develop and change over time.


The key is to personalise each robot for the individual: they cannot be the same for everyone. The aim is for each to be built the same, but to be capable of adaptive behaviour, with a continuous lifelong interactive learning cycle. An owner needs to be able to 'program' their robot through demonstration, not through software coding: It watches and learns. One of the advanced tools used is quantum statistical learning.


Prof. Demiris gave us a whirlwind glimpse into the architecture of their robot applications (HAMMER) and the learning algorithm. For each user, the robot has to learn: How they look; How they move; How they solve problems; How they change over time.

I learnt some of the basic architecture terms, such as: Forward Model ("what" do we want to do); Inverse Model ("How", the controller); Delay Loop (another Forward Model); Feeds to the Muscle; Action Recognition (what did we do). Intriguingly, they have several of those in parallel that provide corrective and confidence-building signals allowing the bad options to be 'killed'.

 

There are many challenges being overcome, such as noise and missing data. Also, adaptive training in high performance scenarios, such as a driving game: at the next play, the game can adapt the challenges specifically to stretch the skill of the player.


Fascinatingly, they are giving a robot interactive memories, so that it can replay what it did some days before (for comparison and showing how the user has progressed).


Such work is not driven by a profit motive, as it is not aimed at the mass market. 

Their robots are intended as a short term measure, while they help the user learn or re-learn how to perform tasks or use equipment better.


At the conclusion of the Q & A, I could only agree that these robotic system do not pose a risk to our society; the greater risk is of our society turning people into robots!


I apologise profusely for undoubted errors in any of my above observations; in my defence, there was such a lot of mind-boggling concepts and techniques that enthralled me and hampered my note-taking.


Don't take my word for it! For more information, papers & videos visit: