Rodney A. Brooks and Maja J. Mataric. Real robots, real learning problems. In Jonathan H. Connell and Sridhar Mahadevan, editors, Robot Learning. Kluwer Academic Publishers, 1993.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....problem because it guarantees action value estimates that asymptotically approach the correct values, independently of the followed action policy. The use of RL in autonomous robot control has been gaining space in the last few years. Some of the main issues involved in real tasks are reported in [14]. The concept of learning based on behaviours instead of low level actuator control was proposed by Mataric in its PhD thesis [15] and later on this approach was adopted by other authors [16, 17] Dorigo and Colombetti combined RL and Classifier Systems to produce highly autonomous robots [18] ....

R. A. Brooks and M. J. Mataric. Real robots, real learning problems. In J. H. Connell and S. Mahadevan, editors, Robot Learning, chapter 8, pages 193--213. Kluwer Academic Publishers, 1993.

....of problems. For simulation purpose time must often be discretized. Also, simulators typically work in an abstract feature space and might therefore ignore key factors for the robot behavior [11] Others argue that simulated controllers are doomed to succeed because of the design of the simulators [3]. As a consequence, software that succeeds in simulation may fail on a real robot [4] Accurate and numeric simulations are typically extremely complex and expensive in terms of computational resources and are thus performed by parallel and distributed simulation [9,13] Despite these problems a ....

R.A. Brooks and M.J. Mataric: Real Robots, Real Learning Problems, in Robot Learning, Jonathan H. Connell and Sridhar Mahadevan, eds., Kluwer Academic Press, 193-213, 1993.

....problem, trying to find solutions for building a computational model which will allow the implementation of the needed hierarchy of behaviour modules. In this paper, we will take another path and try to find a solution based on the evolution of the right hierarchy of action. As hinted by [Brooks 93] the ability for an actionselection process to adapt itself so as to deal with new situations and new problems is a key point for achieving autonomy in an agent. On path to explore is the use of genetic algorithms, and more specifically genetic programming [Koza 92] for evolving controller. ....

R. BROOKS et M. MATARIC Real robots, real learning problems. In J. CONNELL et S. MAHADEVAN (edited by) Robot learning. Kluwer Academic, 1993.

....of e ort is required to implement these systems. Researchers have recognized that an approach with more potential for the development of cooperative control mechanisms is autonomous learning. Hence, much current work is ongoing in the eld of multi agent learning (e.g. 2] Brooks and Mataric [3] identify four types of learning in robotic systems: Learning numerical functions for calibration or parameter adjustment, Learning about the world, Learning to coordinate behaviors, and Learning new behaviors. Our research has examined several of these learning areas. In the rst ....

Rodney A. Brooks and Maja J. Mataric. Real robots, real learning problems. In Jonathan H. Connell and Sridhar Mahadevan, editors, Robot Learning. Kluwer Academic Publishers, 1993.

....a learner a (controller) to be trained and a trainer that provides only a scalar reinforcement signal. Although RL method fits very nicely to robot learning, it is a slow learning process it takes too long for the controller to converge toward the desired performance. There are many reasons [2, 8, 9] that contribute to the slow convergence of RL. The major one, however, is that the controller does not know before hand where to search in action space for suitable reactions. This problem stems from the definition of RL: reinforcement based learning robots learn by doing and a In this paper we ....

Rodney A. Brooks and Maja J. Mat'aric. Real robots, real learning problem. Robot Learning, pages 193--214, 1993.

....n 1=n 1 T r.e. width(T ) n 1=n 1=n T rec. rank(T ) n 1= n 1) 1= max(1; n) T r.e. rank(T ) n 1= n 1) 1= n 1) 6 Conclusion and Future Work We believe the present paper provides hope for escaping from the dilemma in computational learning theory (as well as in work with real robots [8]) that learning is too unsolvable or infeasible. We have provided above some reasonable forms of additional information that yield at least slightly positive solvability results. Future work could investigate improved forms of practically available additional information toward finding ....

R. Brooks, M. Mataric. Real robots, real learning problems. In Robot Learning (Edited by J. Connell and S. Mahadevan), pp. 193--234, Kluwer Academic Publishers, Boston, 1993.

.... It provides a basis for quantifying embodiment, which is significant for behavioural robotics, for example with regard to understanding how to calculate and maximise embodiment, as well as understanding the problems that arise in moving between simulation and actual physical environments (cf. [1]) The ontological neutrality of the definition also enables inter disciplinary discussion about embodiment, for example between the behavioural robotics and intelligent software agents communities. It does this by providing a common framework for addressing embodiment regardless of context ....

Brooks, R.A., Mataric, M.J.: Real robots, real learning problems. In: Connell, J.H., Mahadevan, S. (eds.): Robot Learning. Kluwer (1994) 193-213

.... of building a real robot in a real situation, it is more dynamic and complicated environment than in [ Asada et al. 1994 ] and ffl from a viewpoint of robot learning, existing works have not demonstrated the ability to use previously learned knowledge to speed up the learning of a new policy [ Brooks and Mataric, 1993 ] To the best of our knowledge, only a few works related to the problem have been presented. Whitehead et al. Whitehead et al. 1993 ] proposed a method which learns a multiple goal behavior by decomposing a task into subtasks and merging policies independently obtained in these subtasks ....

....a real situation. The merit of the computer simulation is not only to check the validity of the algorithm but also to save the running cost of the real robot during the learning process. Still, real experiments is necessary because the computer simulation cannot completely simulate the real world [Brooks and Mataric, 1993]. We have done the real experiments for the rst task [Asada et al. 1994] that is, the robot learned how to shoot a ball into the goal without any enemy. Now, we are developing the real experiments for the coordinated behavior. Therefore we show the simulation results for the coordinated ....

R. A. Brooks and M. J. Mataric. \Real robot, real learning problems". In J. H. Connel and S. Mahadevan, editors, Robot Learning, chapter 8. Kluwer Academic Publishers, 1993.

.... adaptive and intelligent behaviors have so far been poorly retributed: whereas building extensive responsive systems has quickly proven to be far too difficult to achieve, giving robots simple learning strategies and letting them figure it out by themself has not produced the expected breakthrough [5]. A more interesting approach would suggest to develop imitative skills for robots so that they can observe and benefit from the experience of others [11, 12] Unfortunately, how to implement such skills on a computer hardware is still an unanswered question. However, we can already argue than ....

R. Brooks and M. Mataric. Real robots, real learning problems. In J. H. Connell and S. Mahadevan, editors, Robot Learning. Boston MA: Klower Academic, 1993.

....this approach can be used to model sensory perception of a mobile robot, as well as to model the behaviour of a specific robot in its target environment. 1 INTRODUCTION The advantages of numerical modelling of robotenvironment interaction are well known and widely appreciated in the literature [2, 3, 4, 11]. Compared with conducting experiments with real robots, simulation is fast, cheap and, perhaps most importantly, facilitates repeated experiments under identical conditions: a property desirable if the influence of certain parameters upon a robot s behaviour is to be determined. However, the ....

....under identical conditions: a property desirable if the influence of certain parameters upon a robot s behaviour is to be determined. However, the disadvantages of numerical models are also known and have so far prevented simulation from having significant impact upon mobile robotics research [2]. Typically, the numerical models of robot environment interaction that are used for simulation are so simplified that their predictions do not match experimental observations well [5, 7, 12, 13] so much so that the outcome of a simulation can only be used for prediction of a real robot s ....

R. A. Brooks, M. J. Mataric, Real Robots, Real Learning Problems. In Chapter 8, Robot Learning, Edited by J. H. Connell and S. Mahadevan, Kluwer Academic Publishers, 1993.

.... particular challenges to learning in terms of the form and amount of information that is available [11, 1] Reinforcement learning, which only requires scalar feedback, has found use in learning policy or action functions (the right set of actions to perform in each sensory state) for robots [11, 8, 10]. Availability of an internal model of the environment often simplifies the learning of action functions [11, 18] Robots that have to navigate and manipulate objects in space can benefit from structures for acquiring and using internal models or spatial maps of their environments. Such models can ....

R. Brooks and M. Mataric. Real robots, real learning problems. In J. Connell and S. Mahadevan, editors, Robot Learning, chapter 8, pages 193--213. Kluwer Academic Publishers, MA, 1993.

.... ffl from a viewpoint of building a real robot in a real situation, it is more dynamic and complicated environment than in [8] and ffl from a viewpoint of robot learning, existing works have not demonstrated the ability to use previously learned knowledge to speed up the learning of a new policy [10]. To the best of our knowledge, only a few works related to the problem have been presented. Whitehead et al. 5] proposed a method which learns a multiple goal behavior by decomposing a task into subtasks and merging policies independently obtained in these subtasks (subgoals) In their scheme, ....

....situation. The merit of the computer simulation is not only to check the validity of the algorithm but also to save the running cost of the real robot during the learning process. However, still real experiments is necessary because the computer simulation cannot completely simulate the real world [10]. We have done the real experiments for the rst task [8] that is, the robot learned how to shoot a ball into the goal without any enemy. Now, we are developing the real experiments for the coordinated behavior. Therefore we show the simulation results of the coordination of multiple behaviors by ....

R. A. Brooks and M. J. Mataric. \Real robot, real learning problems". In J. H. Connel and S. Mahadevan, editors, Robot Learning, chapter 8. Kluwer Academic Publishers, 1993.

....from that of a simulated robot environment system. But is this not mainly a matter of whether the complexity of nature and its laws can be captured by a computer program There have been various discussions on the simulation versus real world debate which should not be reviewed here (e.g. [5]) In my view, the debate itself is not very useful in the sense of trying to find the right answer to the question of what environment to chose for artificial intelligence or artificial life research. Humans are from their early childhood on experts at taking various abstract or fictional ....

R. A. Brooks and M. J. Mataric. Real robots, real learning problems. In Connell J. H. and Mahadevan S., editors, Robot learning, pages 193--213. Kluwer Academic Publishers, 1994.

....itself. Learning has the purpose of facilitating the actions the robot takes, by making them more relevant, appropriate, or precise. The robot s actions are determined by information at different levels. Therefore, different types of learning will be applicable. For example, Brooks and Mataric [6] identified the following types of learning: ffl Learning numerical functions for calibration or parameter adjustment . This type of learning optimises operational parameters in an existing behavioural structure. ffl Learning about the world . This type of learning constructs and alters some ....

....of solving the i th reinforcement learning problem by using the knowledge it acquired from solving earlier learning tasks. 2. 4 Main Methods of Learning The tradeoff between the amount of built in knowledge and learned information has been acknowledged as one of the key issues in robot learning [6]. While reducing the built in knowledge eases the programming task and reduces the learning bias, it slows down the learning process, and restricts therefore the applicability of automatic learning. There is a great variety of useful learning techniques such as neural networks, evolutionary ....

R. A. Brooks and M. J. Mataric. Real robots, real learning problems. In Connell and Mahadevan [10], chapter 8, pages 193--213.

No context found.

R. A. Brooks and M. J. Matari c. Real robots, real learning problems. In Connell and Mahadevan [31], chapter 8, pages 193--213.

....of a learning mobile robot is predicted using our network based approach and shown to produce more faithful results than classical robot simulation methods. 1 Introduction The advantages of numerically modelling robot environment interaction are numerous and widely reported in the literature ([1, 3, 4, 8]) Simulation allows fast and cheap prediction of robot behaviour. It allows simple modifications of the experimental scenario and experimental parameters, and it allows deterministic repetition of experiments, thus facilitating quantitative analysis and replication. However, to obtain these ....

....accurate to exhibit the necessary faithfulness in prediction. Furthermore, there is the real risk that unrealistic simulations may mislead the user, for example by causing him to work on artificial problems. As Brooks and Mataric state, many simulation artifacts do not occur in the real world ([1]) Likewise, there is the risk of missing significant features and therefore basing the design of a control structure on false assumptions. Our previous work, which will be mentioned briefly in this paper, has shown that it is possible to obtain models of robot environment interaction by using ....

R. A. Brooks, M. J. Mataric, Real Robots, Real Learning Problems. In Chapter 8, Robot Learning, Edited by J. H. Connell and S. Mahadevan, Kluwer Academic Publishers, 1993.

....a real situation. The merit of the computer simulation is not only to check the validity of the algorithm but also to save the running cost of the real robot during the learning process. Still real experiments are necessary because the computer simulation cannot completely simulate the real world [15]. 5.1 Simulation We performed the computer simulation with the following speci cations (the unit is an arbitrary scaled length) The eld is a square of which side length is 200. The goal post is located at the center of the side line of the square (see Fig.2) and its height and width are 10 and ....

R. A. Brooks and M. J. Mataric. \Real robot, real learning problems". In J. H. Connel and S. Mahadevan, editors, Robot Learning, chapter 8. Kluwer Academic Publishers, 1993.

....save the running cost of the real robot during the learning process. Further, this policy transfer helps us improve the system by nding bugs in the simulation program and di erence between the simulation and the real robot system. The computer simulation cannot completely simulate the real world [11]. 7.1 Simulation 0 20 40 60 80 100 120 140 0 1 2 3 4 5 6 7 8 9 10 time step M with LEM without LEM Figure 6: Change of the sum of Q values. We performed the computer simulation with the following speci cations (the unit is an arbitrary scaled length) The eld is a square of which side length ....

R. A. Brooks and M. J. Mataric. \Real robot, real learning problems". In J. H. Connel and S. Mahadevan, editors, Robot Learning, chapter 8. Kluwer Academic Publishers, 1993.

No context found.

Rodney A. Brooks and Maja J. Mataric. Real robots, real learning problems. In Jonathan H. Connell and Sridhar Mahadevan, editors, Robot Learning. Kluwer Academic Publishers, 1993.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC