|
|
|
|
Since I started the research of biped robot, I have got many questions from many people. I suppose this is because biped locomotion is one of the most familiar thing that we can do. Some of these questions are ingenuous and some are professional. I believe it is worth to answer the questions in all spectrums between them. I will be happy if you can obtain some hint what biped locomotion is like. But remember, this is only my personal opinion.
Q: Why do you study biped robots ? I think wheels are good enough to move around on a ground. If the robot must go uneven terrain, we can use a caterpillar tractor system likes a tank.
A: One typical example that can not be explored by wheeled robot or caterpillar tractor is step stones in a pond you can see in Japanese garden. A walking machine can utilize "discreet" footholds while wheels and caterpillar need "continuous" support from the ground.
Q: You should make a robot with four or six legs, which is safe from a danger of falling down. Shouldn't you?
A: A biped robot body can be made shorter along the walking direction, hence the robot can turn around in small area. This is useful in an involving environment like in a nuclear power plant. Moreover, biped robot has advantages for light weight and efficiency because the robot needs less numbers of actuator.
Q: How about a walking machine with three legs ?
A: I agree the idea is very attractive. Because it can stand very stable while the locomotion would be dynamic. The problem is it is not known yet suitable walking pattern that deploys three legs equally keeping dynamic stability. If two of the legs are used symmetrically (like R2-D2 in StarWars), we can treat it as a problem of biped locomotion.
Q: A baby learns biped walk after his/her birth. I think it is very useful to study this learning process.
A: Many kind of baby birds and animals can start walking within a few minutes of birth. This suggests us that walking control system of such animals are "genetically programmed". I think the reason that human babies need about ten-month after their birth is they born at too premature status and they do not have enough muscle-skeleton system to start biped walking.
Q: But, I think a child needs to learn walking from their parents. For example, a girl brought up by wolf could not do biped walk but walk with four limbs.
A: That is because neural circuits need appropriate inputs from an environment to reach final goals, which itself are even genetically programmed. This is similar to a kind of computer games called "roll playing game." While all possible results are programmed in advance, the system needs player (i.e. adequate stimulus from outside) to reach the goal (Such style of learning is described in "Society of mind" by Marvin Minsky).
Q: Well, you are believing we are genetically programmed to control our biped locomotion. I understand. Even admitting your hypothesis, it is still true that human had obtained biped skill from the evolution. I feel the use of a genetic algorithm is a promising approach to study biped locomotion.
A: Researchers like Douya and de Garis had challenged to create biped locomotion by simulated evolution and some of their achievements are very fascinating. Mathematically, they are doing numerical exploration in parameter space. As long as treating simple model and limited number of parameters, a numerical exploration works. But, I am wondering if they can still obtain "good result" when the number of parameters becomes huge.
Q: Why don't you expect, in the near future, quick advance of computer technology will make it possible. The complete emulation of human evolution in biped locomotion.
A: All right, even if it becomes possible, that is not enough for me. When we get something by genetic algorithm, we can't answer the question about control system's architecture. For example, "Why it works?" or "Is it necessary and sufficient system for the purpose ?" Only traditional approach can answer to these questions.
Q: Biped locomotion is quite human activity. So, you may use fuzzy controller or neural network controller which was originaly inspired from human information processing.
A: Kawaji et.al successfully realized three dimensional biped walking robot using fuzzy controller, and Kurematsu et.a., proposed biped control system using artificial neural network controller. They carefully analyzed biped walking in traditional way before applying fuzzy or neural technique. Of course fuzzy and neural network controller is very powerful technique for engineering. But, we must recognize they are not almighty.
Q: How about Subsumption Architecture(SA) proposed by Professor Rodney Brooks in MIT ? His six-legged walking robot can walk over various obstacles. Can't we use his theory ?
A: Professor Brooks' thesis was ``Does intelligent systems really need central representation ?" This is very important question to the methodology which current robotics researchers relies on. On the other hand, SA itself is a method of graphical representation for a computer programming, like flow chart or PAD, and it does not give us a magical problem solving power. If necessary, I can rewrite the control program of Meltran II into the subsumption architecture style.
Q: How about understanding biped locomotion as a nonlinear oscillation process? For example, even a cat whose cerebrum was removed can walk on a treadmill. This means a cat has a Central Pattern Generator (CPG) in his spinal cord.
A: Obviously, a biped locomotion is not a linear oscillation, thus it is a non linear oscillation. People working on CPG hypothesis tend to assume very limited nonlinear system like van der Pol equation. From my point of view, the real neural network system of animals might be able to perform far more complicated computation. What the experiment of cat suggest us is, there exists a walking control system in a cat's spinal cord, and also exists a switch which can activate entire walking control. I will not be surprised if the spinal cord does high level computation like inverse kinematics.
Q: You separate everything into smaller parts and expect to understand original by analyzing those parts. But I think it is an old paradigm of science. To understand a complicated system like biped walking, I think the key word will be "holism", which regards total system behavior more.
A: Frankly, I don't see any clear difference between the holism approach and the traditional approach. When a constraint control was applied to a simple nonlinear pendulum, it yields interesting linear dynamics. This is a good example of "The total system is more than a sum of subsystems," but still we can understand it in an analytical way. In my opinion, current Robotics needs more rigorous analytical approach rather than a new paradigm.