Robot, object, action!

Oct 29, 2010
Robot, object, action!

Robotic demonstrators developed by European researchers produce compelling evidence that ‘thinking-by-doing’ is the machine cognition paradigm of the future. Robots act on objects and teach themselves in the process.

European researchers at the PACO-PLUS project have developed robotic demonstrators to illustrate a new approach to robotic cognition. The team uses a concept called ‘object-action complexes’ (OACs), pronounced oaks.

These complexes represent a combination of perception of and action upon any given object. The OACs are recorded by the every time it performs an action. These recorded actions can be saved and exchanged with other robots, and over time the team hope it will lead to entire libraries of OACs that can be exchanged between researchers.

But that is just a major bonus of this particular system. The primary thrust was to show that robots could teach themselves simple actions, leading to more complex actions and, hopefully, leading to abstract thought in the future.

The PACO-PLUS team performed a large number of demonstrations with humanoid robots and the results so far are very promising.

Autonomous exploration

With autonomous exploration, a robot can patiently explore thousands of potential ways to interact with an object it has no prior knowledge of – learning over time that grasping it in one way is better than another, that it balances better in one direction than another, etc.

“The first thing a robot will do when it encounters an object for the first time is to lift it before its eyes and then rotate the object so the robot gets a look at it from all angles. After that, [the robot] will be able to identify the object from any angle,” explains Tamim Asfour, leader of the Humanoids Research Group at the Institute for Anthropomatics at the Karlsruhe Institute of Technology (KIT) in Germany and co-coordinator of the PACO-PLUS project.

One can see how complex actions could arise over time, as a robot progressively teaches itself how to interact with a cup, jug, handles and doors. Similarly, as the robot’s experience with different objects expands, its options expand, too, and it could learn to use a key with a hole in a door to see if that action will lead to any result.

This is the core of the OAC concept, that interaction with the environment leads to ever-more sophisticated strategies over time, and ultimately gives rise to ‘intelligence’.

Of course, the autonomous exploration is extremely time consuming, unsurprising when one recalls that it took hundreds of millions of years to evolve intelligent life. So another method for teaching robots is learning from human observation and human coaching, and PACO-PLUS developed a large range of successful demonstrators with this method as well.

Human coaching

With coaching, robots can watch humans perform an action, for example wiping a table, and then they can imitate it. It is astonishing how easy it is for robots to mimic complex actions like this.

Similarly, a coach can guide the robots through a complex task, for example putting a cup in a dishwasher. At each stage the coach tells the robot ‘open the dishwasher’, ‘take this object/cup’, ‘put it in the dishwasher’. Over time, the robot learns these sorts of actions.

This is a phenomenally sophisticated demonstration of robotic learning, and it was achieved, using a new approach, in the space of just four years. As such it stands as a remarkable testament to the work ethic of the PACO-PLUS team and the power of its approach.

“Of course, we did a lot of other work too,” notes Asfour. “For example, we looked at some situations with more traditional industrial robots, to see how our work could be applied there.”

The work has been very well received in the wider robotic community, and some partners are looking at how they can integrate some of the results into other projects. In the meantime, PACO-PLUS is currently pursuing further research funding to perfect the work achieved.

But already PACO-PLUS has dramatically advanced the field of robotic cognition and helped to set the research and development agenda for the next decade.

This is the second of a two-part PACO-PLUS feature. Part 1.

Explore further: Q&A: Drones might help explain how tornadoes form

Related Stories

Robots use their hands to 'think'

Oct 18, 2010

Action-centred cognition is a groundbreaking concept in robotics where robots learn to 'think' in terms of what actions they can perform on an object. This new trend in cognition theory opens exciting new ...

As robots learn to imitate

Dec 22, 2004

Can robots learn to communicate by studying and imitating humans' gestures? That's what MIRROR's researchers aimed to find out by studying how infants and monkeys learn complex acts such as grasping and transferring it to ...

'What can I, Robot, do with that?'

Apr 21, 2008

A new approach to robotics and artificial intelligence (AI) could lead to a revolution in the field by shifting the focus from what a thing is to how it can be used.

Robots to help children to form relationships

May 29, 2007

A project which is using robots to help children with developmental or cognitive impairments to interact more effectively has just started at the University of Hertfordshire.

Recommended for you

As dust clears, what's next for Sony?

2 hours ago

The Sony hacking attack continues to deliver more dramatic plotlines than any fictional movie, but meanwhile the movie studio must move forward and tackle the next steps in minimizing the mess. Will Sony ...

FAA, industry launch drone safety campaign

5 hours ago

Alarmed by increasing encounters between small drones and manned aircraft, drone industry officials said Monday they are teaming up with the government and model aircraft hobbyists to launch a safety campaign.

It's down to the wire for online shopping

5 hours ago

As the holiday shopping season winds down, FedEx, UPS and online retailers are using the last few days to try to avoid the problems that occurred last year when severe winter weather and a surge in late orders ...

User comments : 9

Adjust slider to filter visible comments by rank

Display comments: newest first

Deadbolt
3 / 5 (2) Oct 29, 2010
There's one problem the robots need to get around.

Let's say I did a load of random movements before putting the cup in the dishwasher, like spinning around and making odd gestures with my hands.

Would the robot be able to make the logical inference that such movements are superfluous to putting a cup in a dishwasher, as a human would, or would it unintelligently copy the movements exactly and associate them with the processes needed to put a cup away.
JamesThomas
5 / 5 (1) Oct 29, 2010
There's one problem the robots need to get around.

Let's say I did a load of random movements before putting the cup in the dishwasher, like spinning around and making odd gestures with my hands.

Would the robot be able to make the logical inference that such movements are superfluous to putting a cup in a dishwasher, as a human would, or would it unintelligently copy the movements exactly and associate them with the processes needed to put a cup away.


My sense is that such programing would have one of its basic algorithms being to figure-out the most efficient way to move object x from point a. to point b.; or calculating efficiency in whatever the task at hand is.
trekgeek1
5 / 5 (2) Oct 29, 2010
There's one problem the robots need to get around.

Let's say I did a load of random movements before putting the cup in the dishwasher, like spinning around and making odd gestures with my hands.

Would the robot be able to make the logical inference that such movements are superfluous to putting a cup in a dishwasher, as a human would, or would it unintelligently copy the movements exactly and associate them with the processes needed to put a cup away.


My sense is that such programing would have one of its basic algorithms being to figure-out the most efficient way to move object x from point a. to point b.; or calculating efficiency in whatever the task at hand is.


Yes, but sometimes style plays a part in society. We have methods that are not efficient but are socially expected. A robot needs to sacrifice efficiency to fit in.
neutrino64
5 / 5 (1) Oct 29, 2010
They mention the slow progression of learning using this trial and error method of interaction. I wonder how effective it would be to use this same methodology, but in a simulated environment instead of a real one. One could take advantage of the natural speed of computers and greatly speed up the process of trial and error.

I've read about this kind of thing in designing structural parts where it is put under simulated loads over and over in an "evolutionary" design process. It seems the same could be applied to this kind of thing. But I'm no computer scientist so perhaps not.
Modernmystic
5 / 5 (2) Oct 29, 2010
If you want true AI don't try to re-invent the wheel. This is very similar to the way human babies learn and I think is the "right track"

You are never going to "program" a true AI in the traditional sense of the word. Any true AI is going to learn, not be programed. Intelligence is a bottom up proposition, not top down.
plasticpower
5 / 5 (3) Oct 29, 2010
My sense is that such programing would have one of its basic algorithms being to figure-out the most efficient way to move object x from point a. to point b.; or calculating efficiency in whatever the task at hand is.


That's assuming the robot understands the purpose of doing that action in the first place. If it's to store a cup in the dishwasher (or wash it) then, as long as it can understand THAT, it should be able to filter out actions that aren't necessary to put away the cup. But if it's just mimicking actions (which sounds like what's happening here) then the robot might think it's a good idea to do that at first (just like a kid does) until it "realizes" the general purpose of this sequence of actions and optimizes its routine.
DamienS
not rated yet Oct 29, 2010
plasticpower, you've said exactly what I was going to. The object here isn't to find the shortest path or most energy efficient way to perform a specific task, but to mimic behaviours in order to expand knowledge of the physical world. With practice, will come efficiency.

This mechanism is used almost universally among higher animals that look after their young, so it has to be a good strategy. The challenge is in making the software general and flexible enough to be able to learn from these behaviours and build on them over time and for the hardware to have sufficient degrees of freedom to mimic coached actions. Nice work.
Modernmystic
not rated yet Nov 01, 2010
Anyone here seen the "Sarah Connor Chronicles"?

It's a little goofy, but there are some interesting philosophical problems brought up in the series.
ShadowRam
not rated yet Nov 01, 2010
This is pointless.

This would only apply if ALL robots were 100% identical. They are not.

Please sign in to add a comment. Registration is free, and takes less than a minute. Read more

Click here to reset your password.
Sign in to get notified via email when new comments are made.