10/VII/2006
Conference Day One: The History of AI and its Goals
The day started at 8:00 having breakfast with all the attendants to the conference. We were impressed by a robot that prepared toasts with butter and marmalade
Having breakfast with VIP of AI
The toast-preparator. An incredible robot only made of mecanics with no electronic control at all. And it works!!
We continued at 8:30 with a talk of Nils Nilson. It was the first of the speakers of a day dedicated to the study of what has been done on AI until now, and what should we expect from it. The first track was called History and Prospects of AI which included two talks (of Nilson and Rodney Brooks), and a panel discussion.
The first speaker was Nils Nilsson, one of the creators of the Shakey Robot. His talk, entitled Routes to the AI summit, was a good description of the path, he thoughts, should guide us to the final achievement of artificial intelligence. Nilsson quickly defined himself as a supporter of Goof Old Fashioned Artifial Intelligence (GOFAI), and he really thinks that this is the path we must follow for achieving the goal of human level AI.
Nils Nilson explaining cognitive substrates for AI
His talk was structured using a how to climb a mountain metaphor. As base champs, he established some of the works made by early researchers: McCulloch Pitts nerve cells and the physical symbol system hypothesis (= formal symbol manipulation + designation and interpretation). All those works allowed him to justify that we need a physical symbol system in order to do intelligent action. Then he showed some early successes using those base champs: Shakey planning robot, Schrdlu blocks manipulator robot, some expert systems, and some works using neural networks.
However, he cleverly noticed that we are still far from the original scope of AI, and that some problems are still there (mainly, IA doesn't know what is doing). So after pointing some criticisms made by other people (paying special attention to philosophers like Searle et al.), and indicating the existent debate between researchers in terms like GOFAI vs connectionism, or procedural vs declarative, he signaled that this debate leaded to new routes to the summit: reactive and behavior-based strategies and phylogenetic (evolutionary) strategies.
He showed then a very interesting graphic plotting the generality vs specifity spectrum, which plots the character of the knowledge about the problem to solve vs the computational cost (on one side) and the cost of obtaining the explicit knowledge (on the other). His argument was to show that for modern AI it is required to be in the middle of that plot, and for this, mechanisms will be needed for general, educable, intelligent behavior.
On another point, he indicated how all previous AI works have helped to have a more solid AI (more technical and mathematically richer) and how all those new techniques have been applied to solve real life problems, and he shows some papers describing those applications.
Finally he describes some paths he would explore if he had time enough. First path, try to build an artificial cognitive substrate that could be used to develop better skills through education, training and experience (he cites as interesting reading the paper A cognitive substrate for human-level intelligence by Nick Cassimatis). Second path he would explore is the creation of computational models of the neo-cortex, expanding the computational metaphor.
His talk ended by proposing that a revision of the pandemonium may be required, and showing a list of current research projects that follow the spirit of the summit, like the Darpa Grand Challenge or the Robocup.
Jürgen Schmidhuber corrects some details to Nilson
His speech, entitled AI: Where from and where to? started with a funny presentation of the history of AI.
Pre-history: some automata of the 18th & 19th century (like Vaucasson's duck).
He pointed out the difference between AI and Cybernetics (the study of all possible behaviors of a machine). Then he continued describing Walter's tortoises, and how non-linearities are the key to their behavior (note: Owen Wilson has done a great research work about those tortoises).
Brook's speech was very vivid as can be felt from the pictures
He continued with the history of AI until he reached the Dartmouth conference (whose anniversary we celebrate now!). According to Brooks, key Dartmouth innovations were:
1- modularity
2- ungrounded symbols
3- search
4- going digital as the mechanism for expression of intelligence.
This last point created a breakthrough in the still at that time un-established field of AI. First it was not accepted but then it became mainstream. This lead to an early selection of problems that they thought were difficult to be solved by a computer (chess, theorem proving,...), since they were problems only solved by the most intelligent human minds. After introducing recursion, AI became search based. Search became the core of AI systems, even-though there was no evidence that general search occurs in humans!!!.
Then he reached the Church-Turing Thesis and questions about if the brain could really be seen as a Turing machine. But he shows that brain has been thought by people to be like the greatest human technology available at the time, and the same situation applies to the mathematical theories used to solve AI.
As a final point of his talk, Brooks showed the changes in the population composition that we are suffering at present, and how the situation will become worst in terms of the number of aged people that will compose the society in the short future. This problem together with other problems like low cost manufacturing labor, decreasing use of immigration for agriculture, rise of aging population requiring more services, and changes in job satisfaction, will pull robotics to solve them. This introduces some immediate challenges for AI. In order to be able to have those required robots, Brooks suggested that they will have to have at least the following capabilities:
- Visual recognition capabilities of a two year old child
- Manual dexterity of a six old child
- Language capabilities of a 4 year old child
- Social sophistication of a 10 years old child.
After Brook's talk, it started a panel discussion entitled The future of AI- Classical or embodied or both?
The panel began discussing the necessity to make AI a science by really understanding what we are doing and leaving the ad-hoc situation that at present dominates the field.
Some ideas from the panel indicated the necessity of understanding by building. Also, embodiment is very relevant, but REAL embodiment, i.e. not simple embodiments like khepera robots. Finally, it was suggested that prediction, is the really final goal of intelligence and that it requires modeling. Not everybody agreed.
The panel discussion lead to lunch time. After lunch, a new block of speeches started under the title of Society at large. First speaker was Hod Lipson.
His speech was about creating machines that can create new machines, solutions, etc on a creavite way, that is why it was entitled Curious and creative machines.
He started saying that we are very good at analysis and prediction, but that synthesis is treated like an art. So his main concern was about: can we teach computers to be creative, to generate new solutions?. How can we achieve it?.
He proposed to solve this question through the use of systematic synthesis using evolution. So he applied evolution to the generation of morphologies, or to the evolution of photonic structures.
He claimed that evolving without any explicit fitness, the result is the evolution of a self replication system. And as a prove, he showed the self-replicating cubes. He says that all these are samples of creativity.
Then he moved to curiosity, and asserted that curiosity requires synthesis. He said that curiosity is about asking the right questions, and proposed a model of curiosity based on Schmidhuber's ones. It applies that model for the generation of a self-model (i.e. a model of itself), which leaded to an emergent self-model.
Pfeifer asks a question to Lipson at the end of his talk
His claim was then that modeling is required for asking good questions, but that models must be learned. Furthermore, once we achieve to have a good synthesis system, we obtain that analysis + synthesis = adaptation.
Then he went on the description of a system that is capable of doing the correct questions.
Finally, he concluded that embodiment provides an endlessly-rich substrate for open ended generation of answers and creativity. Not having that embodiment is a great disadvantage. Because of that, he is very concerned to the use of robot bodies. In order to achieve an easy way of creating bodies, he presented a printing system that is capable of printing materials for robotics bodies including thermoplastic, motors and even living cells.
She is the creator of the Kismet robot, and works at present at MIT. She considers home and everyday life as the final frontier that robots must conquer. That is why her talk was entitled Socially intelligent robots.
According to her, robots should be considered not as tools but as social actors. Humans are hardwired to interpret the behavior of others in terms of mental states. Because of that, robots elicit social interpretation and evoke human social responses (if there is a robot that watches you, then you think there is someno watching you). So people bond with robots.
Furthermore, social relation is seen as an interface between humans, but it can reach total socio-emotional and psychological aspects of people. Then human-robot interaction is synergy of the human-robot system.
Following this line of reasoning, Breazeal thinks that the interesting challenge is toward robots that treat people as people, not like objects, and actively relate to people in psychological terms. Because of that she created a robot called Leo which she uses for social robotics experiments.
Cynthia Breazeal
She identified the following cornerstones of social robotics:
- Interdependence
- Transparent communication
- Cognitive compatibility
- Perspective taking
And then focussed on the explanation of some key experiments she's doing for achieving those capabilities:
* Mental perspective: she defined it as how to infer from non-verbal situations how the human partners of the robot feel or mean. She explained some neurological data that suggest that we are using our own cognitive system to simulate others. She paid attention to the capacity to imitate others and make a model of them by using mirror neurons.
She has applied all this to generate a robot (called Leo) than can imitate a human babbling. Then she applies this as intrumental support, by providing helpful actions to the human.
* Leaning from ambiguous demonstrations: it is found in several tests that people adopt demonstrator's perspective when learning concepts. She tries to apply this to her Leo robot.
* Communication for joint action in teamwork: communications are necessary in teamwork in order to align mental states, and to ground mutual beliefs. She has identified some verbal and nonverbal cues for grounding and coordinates activity. She applies it to the Leo robot. She suggests that people will interpret much of the robot behavior as communicating its mental states in a similar way as humans do.
* Communication for joint action in social learning: it is about robots that can learn in the wild. How to build robots that can learn their own learning problems. Can this be socially guidable?. One of the premises is to see tutelage as a transparent collaborative process. Teacher guides exploration, but learner guides the teaching process. She identified intrinsic and extrinsic factors in robots that learn from people.
In the future, she thinks, robots will become common place. People will be attached to them. They will collaborate with people as patners, and will need socio-emotive dimensions for a wide variety of tasks. New applications for robots in society will be found. How will this situation be reached?. She suggested that by understanding coupled human-robot system as a whole. By understanding how human social intelligence works, and how it supports collaborative activities.
Her speech entitled Archipielago.ch: the dynamic diorama, was the strangest of the whole conference. So is the case that I'm not sure if I really undertood what she was pointing out.
As far as I understood, her research explores the function of narrative as a sub-field of creative robotics. She claims that the creation of narrative scenarios in parallel with research in artificial intelligence could be a useful way of introducing a layering of context and meaning which could help to improve both the AI research and the art.
In order to explain her claims, she showed some videos created by herself where she showed the union of artificial intelligence and narrative on a kind of robotic psychoanalizer. Another video showed a strange robot that is running on a New York museum. The robot is autonomous but can be controlled from the WWW by external people to see what happens and interact with the visitors.
The second panel discussion of the day was entitled Advertizing AI to the public and to companies - stratedies and methods.
Discussion on the 90% that remains for the generation of a product after the paper has been created by the researcher
The robot demo was the most exciting event of the whole conference. It took place on the same room where the posters were later presented, and congregated a lot of people, including the local press. This robot demo was some days later repeated for non-professional people.
The robot demo showed robots of all kinds. Basically, it was though as a way of showing AI techniques applied to robotics.
On humanoid robots, we found two different types:
* The first one was an upper torso robot called CRONOS, which could move its arms and hands in order to grasp things or shake hands. The idea behind the project is to study in which way the muscles are connected to the skeleton and act on it to produce movements.
CRONOS robot shaking hands
* The second was the winner of the Robocup 2006 in the humanoid league
The humanoid stands up and starts walking
A close-up of the humanoid standing up
On the field of robots for the study of locomotion, there were four different robots dedicated to it:
* First was the puppy II robot, a robot developed at Zurich that takes into account morphological computation in order to produce a cheap design
Walking style of Puppy robot
* There was that strange robot called Bio Leg (I and II) that appeared in different configurations of one or two legs (the video only shows one leg locomotion). The goal is to identify general principles of legged locomotion and apply them to therapeutics.
One leg walking robot
* As a strange robot too, it appeared the Quad robot, which is a robot with the shape of a sea star with only four legs. This robot is used for the study of different forms of locomotion through evolutionary mechanisms. It is a shame that the batteries did not last long enough to take a video of it walking.
* Finally there was a fish robot from the University of Zurich, used to study the intrinsic principles in locomotion, which leaded to so many different types of locomotion, like swiming or legged locomotion.
Fish robot swimming
On autonomous robots, there was the roomba robot and the Aibo robot. Nothing to say about those two robots, since everybody knows them. On autonomous robot, there was also a flying robot which flight autonomously while avoiding walls, ceiling and people around.
Roomba robot on a crowded room
The flying robot of the EPFL
The Aibo robot trying to learn how to walk taking as reference the pink ball
As an additional autonomous robot, we found an attempt to create an autonomous suitcase, but even that the project has already achieved a lot (as can be seen in the video), its degree of autonomy is still low.
The suitcase travels by itself and avoids an obstacle on the side
Under the section of social robotics, we found the Paro robot. This is a robot with the shape of a baby seal, created as a therapeutic robot. Impressive!!. This robot has already won a lot of international prizes and has also been a great sale success in Japan.
some movements of the Paro Robot
Another section in the exposition was that of robots controlled by the movements of a person using the electric signals of his body. There three interesting robots on this track:
-The first one was the Pet Bot, a robot made of PET bottles of water, that when connected to a human arm, it can interpret the neural signals of the arm as movements of its bottles. The result, a group of bottles of water that walk controlled by the neural signals of a human arm. Eventhough this was a very impressive result, the main goal of this project was to focus attention on the possibilities we have when we want to transfer strange evolved creatures from simulation to real embodiment.
-The second was an artificial hand that could be controlled by the nervous system of a human. Current development centers on introducing feedback or feeling of the movements to the human.
- Third was the ZAR5 robot. A robot based on liquid muscles that can be controlled through the use of a special suit. Due to the use of those bio-inspired liquid muscles, the movements of the robot are smoother and pleasant than in other robots
Other robots included a kind of small industrial robot called Katana
and a set of small autonomous robots that exhibit auto-organization properties when let them free to act with simple vibration movements on water.
Small robots in water, in the process of auto-organization
Note: it may happen that the text contain errors. Please contact me if you find any error, since I will be able to fix it.
