O. Miglino, H. Hautop Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....performed. 4 The Simulator A central issue in any evolutionary robotics systems is the evaluation of candidate controllers during evolution. It was decided to perform the evaluations in a simulation modelling the Khperea soccer world for a number of practical reasons. Using a table based model [MNL95] was rejected due to the large state space of the world. Instead a hybrid between the minimal simulation approach, suggested by Nick Jacobi [Jac98a] and a geometric model is used. Basically ev erything that easily could be modeled geometrical is modeled geometrically, and then the rest is made ....

Orazio Miglino, Stefano Nolfi, and Henrik Hautop Lund. Evolving mobile robots in simulated and real environments. Artificial Life 2, pages 417434, 1995.

....in the real world, such as the walls having specific colors or the robot having perfect wheel odometry, or possibly even bugs in the simulator. This becomes evident when a controller evolved in simulation is transferred to the real robot. This step is referred to as crossing the reality gap [10] [4] How how do we make controllers cross the reality gap Forces: One must consider the complexity of the environemnt, Discreteness of robot and environment, reliability of sensors and actuators, performance fluctuations due to varying battery levels, varying light conditions. In short: all ....

Orazio Miglino, Henrik Hautop Lund, and Stefano Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life 2, pages 417-434, 1995.

....Robotics is an emerging field that uses simulated evolution to produce control programs for robots. Recent work can be found in [7] and [23] Most work uses bitstring Genetic Algorithms to evolve neural nets for obstacle avoidance and wall following using sonar, proximity or light sensors, e.g. [8, 20, 19]. Significantly, gradient based learning techniques consider recurrent neural networks much harder to train than feed forward networks, since gradient information typically isn t available. Evolutionary Computation doesn t use gradient information, and therefore even exploratory, toy problems can ....

O. Miglino, H. H. Lund and S. Nolfi, Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2, 417-434 (1995).

....and [20] Unless otherwise mentioned, the work reported here has been validated by running it on a real robot, not just in simulation. Most work uses bitstring Genetic Algorithms to evolve recurrent neural nets for obstacle avoidance and wall following using sonar, proximity or light sensors, e.g. [8, 17, 16]. Significantly, recurrent neural networks are considered much harder to train than feed forward networks, since gradient information typically isn t available. Evolutionary Computation doesn t use gradient information, and therefore even exploratory, toy problems use recurrence. Nordin et al. ....

O. Miglino, H. H. Lund and S. Nolfi, Evolving Mobile Robots in Simulated and Real Environments. Artificial Life , 2, 417-434 (1995).

....even a small deviation can lead to failure. The accuracy of the world model is an important factor for the success. Most of the work in GP use the simulated world to perform learning. The problem of transferring the result from the simulated world to the real world has been widely recognised [1,2, 5, 7, 10, 11]. To cope with changes, many researchers suggest the use of physical robots to learn in the actual environment of the tasks [5] The robot will learn by trial and error. This approach is suitable for many learning tasks such as learning the association between sensing and effectors. However, the ....

....learn the task. It is possible to reduce the time by running GP in simulation that samples data from the real world [9, 11, 13] Another approach to cope with changes is to subject the evolved system to perturbation expecting that the resulting solution will be more tolerant. The work such as [7, 12] introduce perturbation at every step of evolutionary process. They report limited success. From the experience of our previous work, we notice that we can introduce limited perturbation in between generation with good results, i.e. during a generation, we keep everything constant. Another ....

Miglino, O., Lund, H. and Nolfi, S. , "Evolving mobile robots in simulated and real environments", in Artificial Life 2(4), 1996.

....references to every journal article included in this bibliography. The list is arranged in alphabetical order by the name of the journal. Adaptive Behavior, 86, 237] Advanced Technology for Developers, 414] Artif. Intell. Eng. UK) 218, 233] Artificial Intelligence, 119] Artificial Life, [83, 147, 197] BioSystems, 351] Comput. Ind. Eng. UK) 229] Control Engineering Practice, 90] IEE Colloq. Dig. 262] IEE Conf. Publ. ETSI konferenssi, 265] IEEE Transactions on Evolutionary Computation, 301] IEEE Transactions on Industrial Electronics, 244] IEEE Transactions on Systems, Man, and ....

....Leung, T. P. 400] Lewis, M. Anthony, 401] Li, G. 317] Liegeois, A. 177, 185] Lin, C. S. 303] Lin, Fang Chang, 214] Lin, Hoi Shan, 73, 108, 406] Logan, Brian, 230] Lopez, Luis R. 413] Lott, Christopher G. 81] Louis, Sushil John, 317] Luk, B. L. 265] Lund, Henrik Hautop, [147, 178, 197, 302, 320] Luong, L. H. S. 30] Lybanon, M. 41] Magdalena, Luis, 82] Mantere, Timo, 10] Masayuki, Inaba, 234] Matari c, Maja, 231] Matsunaga, Y. 250] Mazer, Emmanuel, 160, 416, 417, 418, 419, 420, 421, 422, 423, 424] McClain, Jeffrey J. 242] McDonnell, John R. 402, 403, 404, 405] ....

[Article contains additional citation context not shown here]

Orazio Miglino, Henrik Hautop Lund, and Stefano Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, Summer 1995. ga95aMiglino.

....in simulation that have transferred to real robots: locomotion for a hexapod [Gallagher Beer, 1992] and [Gallagher et al. 1996] obstacle avoidance with a Khepera [Michel, 1995] and others. One approach to developing a simulator is to create a simulator based on data taken from a real robot, [Miglino et al. 1995] and [Lund Miglino, 1996] Actual robot sensor readings are used to create lookup tables. In their experiments they evolved a neural control system for a Khepera to move around a 4 AIBO is a registered trademark of Sony Corporation simple environment while avoiding obstacles. Limitations of ....

Miglino, O., Lund, H., & Nolfi, S. (1995). Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434.

....the sensor simulation method used by Michel has been replaced by a tabulation technique, according to which, prior to evolution, the values returned by a sensor in a given environment are recorded in a look up table for a number of different positions and orientations. Note that unlike in [30], where the values stored were measured on the real robot, here, we synthetize these values to make it easier to change the environmental conditions. At evaluation time, the sensor values are computed by interpolation from the values stored in the table. Finally, another important feature to ....

....robot then avoids the region where the lamp is situated and where such disturbing light reflections are the strongest. 4 Discussion Although obstacle avoidance would appear a behavior easy to evolve in Khepera, as demonstrated by the successful results already obtained by numerous researchers [6,9,8,17,24,28 30,34], results presented herein indicate that such a behavior is easily disrupted when the ambient light is high. These results also indicate that evolving a robust obstacle avoidance behavior, although not trivial, is nevertheless possible. The solution that has been automatically discovered here ....

O. Miglino, H. Lund and S. Nolfi, Evolving mobile robots in simulated and real environments, Artificial Life 2 (1995) 417--434

.... (Ziemke, 1999) which documents detailed comparisons between first and higher order recurrent networks in two experimental setups (using an evolutionary algorithm) In the case discussed here, as illustrated in Figure 9, a simulated Khepera robot (using a modified version of the simulator of Miglino et al. 1995)) was placed in an environment which contained 11 identical round objects, five of them inside a circular zone (surrounded by a borderline) and six of them outside. The robot controller receives input from four infrared proximity sensors at the front and a simulated ground sensor which was ....

Miglino, Orazio, Lund, Henrik H., and Nolfi, Stefano (1995). Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417-432.

....the sensor simulation method used by Michel has 3 been replaced by a tabulation technique, according to which, prior to evolution, the values returned by a sensor in a given environment are recorded in a lookup table for a number of di erent positions and orientations. Note that unlike in [30], where the values stored were measured on the real robot, here, we synthetize these values to make it easier to change the environmental conditions. At evaluation time, the sensor values are computed by interpolation from the values stored in the table. Finally, another important feature to ....

....the robot then avoids the region where the lamp is situated and where such disturbing light re ections are the strongest. 4 Discussion Although obstacle avoidance would appear a behavior easy to evolve in Khepera, as demonstrated by the successful results already obtained by numerous researchers [6, 9, 8, 17, 24, 28, 29, 30, 34], results presented herein indicate that such a behavior is easily disrupted when the ambient light is high. These results also indicate that evolving a robust obstacle avoidance behavior, although not trivial, is nevertheless possible. The solution that has been automatically discovered here ....

O. Miglino, H. Lund and S. Nol, Evolving mobile robots in simulated and real environments, Articial Life 2 (1995) 417-434

....wall following or target finding, under the selective pressure that dedicated fitness functions afford. For instance, to evolve the controller of a Khepera robot moving and avoiding obstacles in a given environment, the following fitness function with three components is used in [9] [44], 36] 58] F = V: 1 Gamma p D) 1 Gamma I) 1) where V is the sum of the wheel speeds at each time step, D is the signed sum of the absolute differences between the speeds of the two wheels at each time step, and I is the sum of the largest of the eight infra red proximity sensor values at ....

....[20] neurons of intermediate complexity are used, which propagate excitatory and veto signals to other units after specific timedelays associated with each connection. The architectures of the neural controllers that have been evolved to control robots range from simple perceptrons (e.g. [44], 36] 58] to partially recurrent Elman like [8] networks (e.g. 9] 45] to fully recurrent continuous time (e.g. 66] 17] or discrete time (e.g. 4] 19] networks. The use of recurrent connections affords the possibility of managing an internal memory, as mentioned above (e.g. ....

Miglino, O., Lund, H.H. and Nolfi, S. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life. 2, 417-434. 1995.

....(e.g. different light conditions) changes in sensor response, re arrangement of environment configuration, transfer from simulated to physical robots, and transfer across different robotic platforms. Some authors have suggested to improve the robustness of evolved systems by adding noise (Miglino, Lund, Nolfi, 1996; Jakobi, 1997) and by evaluating fitness values in several different environments (Thompson, 1998) However, both techniques imply that one knows in advance what makes the evolved solution brittle in the face of future changes in order to choose a suitable type of noise and of environmental ....

....to the neural controller. After 500 sensory motor cycles, the light is switched off and the robot is repositioned by applying random speeds to the wheels for 5 seconds. The experiments have been carried out in simulations sampling sensor activation and adding 5 uniform noise to these values (Miglino et al. 1996). 4 The fitness results reported in figure 4 show that individuals with adaptive synapses and Node Encoding (graph on the left) are much better than individuals with genetically determined synapses and Synapse Encoding (graph in the center) in that: 1. Both the fitness of the best individuals ....

Miglino, O., Lund, H. H., & Nolfi, S. (1996). Evolving Mobile Robots in Simulated and Real Environments.

.... hands can be opened and closed) The gripper also has a light barrier object sensor, which detects whether or not there is an object between the hands. For training of the networks Nolfi used a simulator based on measurements obtained from a real Khepera robot, presented by Miglino et al. in [10]. To validate the evolved controllers, they were downloaded on the real Khepera robot and tested in a physical environment. The results confirmed that the simulator is sufficiently realistic to enable a successful transfer of controllers from simulation to the physical robot. The experiments ....

O. Miglino, H. H. Lund, and S. Nolfi, Evolving mobile robots in simulated and real environments, Artificial Life, 2(4), pp. 417 - 434, 1995.

....Robotics, where, albeit in relatively constrained environments, control systems are developed in a simulation environment, and transferred directly and successfully to robots operating in The Real World. See, for example, Mautner Belew (1999b) Lund Miglino (1998) Wilson et al. 1997) Miglino et al. 1995). Such cases demonstrate that the central operational principles associated with embodiment apply regardless of the ontological type of the domain in which they do so. A core axiom here might be expressed as follows: where behaviour emerges from the interplay between system and environment, if ....

Miglino O., Nafasi K., et al. (1995). Evolving mobile robots in simulated and real environments. In: Artificial Life 2(4): 179-197.

....a simulation is, the longer it takes to run each individual. This leads to the question: how accurate must a simulation be to enable an agent that is designed by a simulation to be useful in a real robot Other researchers who have considered this issue include Beers [1] Jakobi [2] Miglino [4], and Nolfi [5] Each of these researchers has evolved controllers in simulations and then gone on to place the evolved controllers into real robots. Transferring results from simulation to real robots is of critical importance in our research. A simulation that successfully transfers to a real ....

....must include a level of noise injected into those features that provide ranges of values in the real world. This noise must vary between runs but not within runs. In this way robustness to a noisy environment will be evolved into the controller. While Jakobi looks for minimal simulations Miglino [4] used actual sensor and effector values of the robot as the source for their simulator. This approach is thus valid for only one robot in only one environmental setting. Raising the temperature or lowering the light may change the results of such a simulation significantly. In this work we use ....

[Article contains additional citation context not shown here]

O. Miglino, H. H. Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, 1995.

....atten tion must be paid to the model as its accuracy directly effects the results. The time and effort can sometimes exceed the work required to program by hand. If most of the training is done off line and then transferred to the actual robot for some remaining generations [Lund 1996, Miglino 1995], then a less accurate model is required, but it can take significant time to do the on line training on the actual robot. If the task can be completed and the fitness can be accurately judged in minimal time, all of the training can be done on line [Husbands 1997, Mondada 1995] This method ....

Miglino, O., Lund, H., and Nolfi S. (1995). "Evolving Mobile Robots in Simulated and Real Environments." Technical Report, Institute of Psychology, C.N.R., Rome.

....will transfer into the real world, and two sets of experiments are briefly described in which controllers that evolved in extremely minimal simulations were able to perform non trivial and robust behaviours when downloaded onto real robots. 1 Introduction Several experimenters including [6, 9, 2, 10] have shown that it is possible to evolve control architectures in simulation for a real robot. Now this is no longer in doubt the question becomes one of whether the technique will scale up. As Mataric and Cliff point out in [8] if robot controllers evolved in simulation can only be guaranteed ....

O. Miglino, H.H. Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artifical Life, 2(4), 1995.

....of the best individual throughout generations. with generations. First we observe that the number of training steps (nbr. of generations 3 nbr. of individuals 3 nbr. of motions) is a factor of 10 to 100 less than the number of training steps used by evolution algorithms applied to neural networks [19] or to production rules [5] Second we observe two types of jumps during the evolution due respectively to mutation and cross over operators. The mutation operator creates individuals avoiding obstacles and generates a large jump of the fitness with a small jump of the number of steps. For example ....

Orazio Miglino, Henrik Hautop Lund, and Stefano Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, summer 1995.

.... of autonomous agent have proposed to use artificial neural networks of some variety as the basic building blocks for the control system of the robot due to its generally smoother search space and its working with very low primitives avoiding using preconceptions about the properties of the systems [14]. Finally Luc Steels proposes to use a dynamical systems, called a process network, inspired by the couple map latticed to control the robot behaviors [20] Boolean Function Controller. In our approach, we assume that the robot behavior is described by a boolean function. It is more simple than ....

O. Miglino, H.H. Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, summer 1995.

....charger under construction. When the robot happened to be over the black area, its simulated battery became instantaneously recharged. A multilayer perceptron of continuous sigmoid units was used to map sensor inputs into motor outputs. 12 input units clamped to 8 infrared sensors, 2 1 But see [16] for a clever methodology that bridges the gap between simulation and implementation in certain conditions. Khepera.eps 56 Theta 40 mm env.picture.eps 50 Theta 67 mm a) b) Figure 2: a) Khepera, the miniature mobile robot. b) The environment with the light tower and the robot (there were no ....

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 1996. in press.

....use of simulation permits the incorporation of specifiable amounts of noise 1 , and the condition of transference ensures that evolutionary robotics remains faithful to real robots in the real world. This approach, already shown to be viable (for example see Jakobi [4] Nolfi [9] Miglino et al. [8]) stands in contrast to two major alternatives. The first, evolution in real time on real robots (e.g. Floreano and Mondada [2] does not easily allow for the explicit, quantitative specification of noise levels 2 and is also formidably time intensive. The second, evolution in simulation with ....

....outlined in section 1, evolutionary robotics has, to date, concentrated on the discovery that noise in simulation can help bridge the reality gap . Controllers evolved in simulations with appropriate noise levels transfer to real robots and real situations (see e.g. Jakobi [5] Miglino et al. [8]) Jakobi ( 5] 4] has formalised the use of noise for facilitating transference by distinguishing between base set features (those aspects of an agent environment system that may come to play a part in the eventual behaviour) and implementation features (those aspects which are either ....

[Article contains additional citation context not shown here]

O. Miglino, H.H. Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, 1996.

.... Signal Sensors Figure 2: Reactive Navigation System networks of some variety as the basic building blocks for the control system of the robot due to its generally smoother search space and its working with very low primitives avoiding using preconceptions about the properties of the systems [22][15] Finally Luc Steels proposes to use a dynamical systems, called a process network, inspired by the couple map latticed developed by Kaneko and various coworkers to control the robot behaviors [29] 17] 3.1 Boolean Function Controller. In our approach, we assume that the robot behavior is ....

....Indeed by generalization it can extrapolate what it has seen so far to similar situations which may arise in the future. To increase the robustness of navigation systems, some researchers either include noise in the environment either use a particular way to evaluate the fitness [15] 13] [22]. In our approach we guide the evolution during the learning process to obtain robustness. To build robust evolutionary controller at one side we force the robot, during its evolution, to encounter many different situations. On the other side we use the generalization capacity of the evolvable ....

O. Miglino, H.H. Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, summer 1995.

....as they moved against similarly patterned backgrounds in a random fashion. 1. Introduction Various different types of simulation have been used in the past to evolve controllers for robots: in (Jakobi et al. 1995) an empirically verified model of the underlying physics was constructed; in (Miglino et al. 1995), look up tables compiled from real world sensor data were used; and in (Yamanuchi and Beer, 1994) the factorybuilt simulation came supplied with the robot. However, the fundamental approach underlying the use of all of these simulations is the same: the less differences there are between ....

Miglino, O., Lund, H., and Nolfi, S. (1995). Evolving mobile robots in simulated and real environments. Artifical Life, 2(4).

....real world evaluation approach advocates. Clearly the alternative approach involving the evaluation of controllers in simulation is preferable: evaluations can be performed at faster than real time, and all of the problems stated above are easily avoided. As has been shown by several experimenters [20, 3, 23], it is possible to evolve controllers in simulation for a real robot. Now that this is no longer in doubt the question becomes one of whether the technique will scale up. In [22] and similar points were made earlier in [14, 6, 10] the authors argue that if behavioural transference can only be ....

....a controller s ability to perform the same behaviour in two different environments. There are several different ways of judging whether a controller successfully transfers into reality after being evolved in a simulation, and the authors that have written about it so far use different methods. In [23], for instance, the authors look at the fitnesses of controllers in simulation and compare them to the fitnesses of controllers when downloaded into reality: the nearer the fitnesses the better the transfer. In [20] on the other hand, the authors use a more subjective approach to judge whether ....

[Article contains additional citation context not shown here]

O. Miglino, H.H. Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artifical Life, 2(4), 1995.

No context found.

O. Miglino, H. Hautop Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 1996.

....to take into account. For instance, the actuators will produce friction and there will be a whole range of noise issues that makes it very difficult to transfer an idealised model from simulation to the real world. In some cases, it will be possible to transfer models from simulation to reality [8, 6, 2]. This is done by very careful building of a simulator that models the important characteristics of the real device and the way that real world noise interferes with this device. In our emergent behaviour model, we work directly in the real world, so we avoid the problems of difficulties in ....

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2(4):417-434, 1995.

....function, but the user performs the selection in the genetic algorithm. In order to use the user guided evolutionary robotics approach, it is necessary to simulate the robot in its environment, make selective reproduction in the simulator, and then transfer to the physical robot. As described in [12, 18], it is possible to build an accurate simulator that allows very good transfer from simulation to reality by basing the simulator on the robot s own samplings of sensor and motor responses. The disadvantage is that data has to be collected. In the construction of the simulator, this data had to be ....

....response for each individual LEGO robot design that we wanted to use in the simulator. This is the disadvantage of the approach. However, we are currently exploring new adaptive techniques to overcome this problem. The sensor and motor data was collected in a similar way to that described in [12, 18], and the collected data was put into look up tables that is used by the simulator to look up specific sensory readings and displacements of the simulated LEGO robot. Our first experiments showed that we could develop simple robot behaviours such as obstacle avoidance, line following, etc. for ....

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2(4):417-434, 1996.

....robot control architecture used in [7] cf. Figure 1) can serve as the basis for internal simulation of perception. 2. Experiments Robot and Environment The experiments documented here have been carried out with a Khepera robot [19] depicted in Figure 4, or to be exact, with a simulator [20] based on sensor and motor measurements obtained from a real Khepera robot. Figure 4: a) Khepera robot built by K Team SA (www.k team.com) b) Schematic drawing of the robot with infrared proximity sensors (1 8) left and right wheel (controlled by independent motors) The robot s diameter ....

Miglino, O., Lund, H. H. & Nolfi, S. (1995). Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2(4), 417.

....the presence of an object between the two arms of the gripper. The environment is a rectangular arena surrounded by walls containing 5 target cylindrical objects, which are positioned randomly inside the arena. The evolutionary process is conducted only in simulation in order to speed it up (Miglino, Lund and Nolfi, 1995). We compare the results obtained with modular and nonmodular neural network architectures (see Figure 2) In both cases the robot has 7 sensor neurons and 4 motor neurons. The first 6 sensory neurons are used to encode the Figure 2. Architectures (a) and (b) are shown on the left and right side, ....

Miglino, O., Lund, H. H., and Nolfi, S. 1995. Evolving mobile robots in simulated and real environments. Artificial Life 4:417-434.

....have had a strong impact on basic research, inspiring a broad range of different studies. Using a special kit of programmable bricks, David Hogg, Fred Martin, and Mitchell Resnick from MIT Media Lab have constructed the main parts of Braitenberg s vehicles (Hogg, Martin, Resnick, 1991) Lund and Miglino (1995) have produced the same series of Braitenberg vehicles using the basic hardware structure presented in figure 3. Figure 3. General configuration of the hardware of a Braitenberg vehicle. Timid (shadow seeker) The robot has one sensor that senses light intensity. The vehicle moves forward if ....

....artificial organisms. For instance, additional work aimed at improving the realism of the artificial cricket was done using an extended version of the Khepera robot, mainly because it is small and offers improved control of experimental parameters. Figure 12. The Khepera miniature mobile robot. Miglino, Lund, and Nolfi (1995) have further elaborated on the UCLA approach using Khepera. At the Italian National Research Council s Institute of Psychology in Rome, the three researchers have used simulated evolution to develop neural network control systems for the Khepera robot and have then gone on to transfer the best ....

Miglino, O., Lund, H. H., & Nolfi, S. (1995). Evolving Mobile Robots in Simulated and Real Environments. Artificial Life 2(4), 417-434.

....have to take into account. For instance, the actuators will produce friction and there will be a whole range of noise issues that makes it very di cult to transfer an idealised model from simulation to the real world. In some cases, it will be possible to transfer models from simulation to reality [8, 6, 2]. This is done by very careful building of a simulator that models the important characteristics of the real device and the way that real world noise interferes with this device. In our emergent behaviour model, we work directly in the real world, so we avoid the problems of di culties in transfer ....

O. Miglino, H. H. Lund, and S. Nol. Evolving Mobile Robots in Simulated and Real Environments. Articial Life, 2(4):417-434, 1995.

....to take into account. For instance, the actuators will produce friction and there will be a whole range of noise issues that makes it very difficult to transfer an idealised model from simulation to the real world. In some cases, it will be possible to transfer models from simulation to reality [Miglino et al. 1995; Lund and Miglino, 1996; Jakobi, 1998] This is done by very careful building of a simulator that models the important characteristics of the real device and the way that real world noise interferes with this device. In our emergent behaviour model, we work directly in the real world, so we ....

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life 2(4), 417-434, 1995.

....alternative. Some aspects of future work are important. First of all, because our work is done in simulation, it is necessary to build a real robot (Lego like) and download the evolved controller to it to observe the performance. Although there are gaps between simulated and real worlds, research [7] has shown that we could sample the real sensor data and the robot motion in the real world to build a more realistic simulator to develop evolutionary systems. Our future work also involves integrating these considerations into our simulator. Finally, since our experiment is focused on a certain ....

O. Miglino, H. H. Lund, S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. To appear in Artificial Life, 1996.

No context found.

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2(4):417--434, 1996.

....each bit with a constant of probability pm=0.02 3 . For each set of experiments we ran 10 replications starting with different randomly assigned genotypes. In this paper we will refer to data obtained in simulation. A simulator developed and extensively tested on Khepera by some of us was used [15]. 2.2 Measuring adaptive progress in co evolving populations In competitive co evolution the reproduction probability of an organism with certain traits can be modified by the competitors, that is, changes in one species affect the reproductive value of specific trait combinations in the other ....

Miglino, O., Lund, H. H., & Nolfi, S. (1995). Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 4(2), 417-434.

No context found.

O. Miglino, H. H. Lund, S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. In Artificial Life, 2(4), 1996.

....to test different experimental conditions and several replications with different random initializations, and to carry out explorations of the fitness landscape. In order to ensure a good match between simulations and physical experiments, we used a sampling technique proposed by Miglino et al. [24]. Each robot was positioned close to a wall of the environment and performed a full rotation by steps of 10 ffi . At every step, all sensor values were recorded and stored in a table. The robot was then positioned 2 mm from the wall and the same procedure was applied again. This technique was ....

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2:417--434, 1996.

....line corresponding to frontal direction) in the arena. For each competition, the initial orientation is random. simulator and initial results are given in (Floreano and Nolfi, 1997b) Here it is sufficient to mention that the simulator was based on real sensory values sampled from the two robots (Miglino et al. 1996), not on a mathematical model of the environment. 3. Results An exploratory set of experiments were performed in simulation to understand the influence of various parameters, such as the number of tournaments with opponents from previous generations, crossover and mutation probabilities, ....

Miglino, O., Lund, H. H., and Nolfi, S. (1996). Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2:417--434.

....to frontal direction) in the arena. For each competition, the initial orientation is random. one can have small discrepancies between behaviors in simulation and on the real robot by sampling sensor activity at different distances and angles of the robot from the objects of the world (see [10] for details) We have thus employed the same methodology and sampled infrared sensor activity of each robot in front of a wall and in front of another robot. These values were then separately stored away and accessed through a look up table depending on the faced object. Simulation of the visual ....

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2:417--434, 1996.

....14] In the field of Evolutionary Robotics, there is much discussion about what kind of robot controller to evolve. Some researchers have developed controllers based on explicit programs in high level language [3, 17] implementations of classifier systems [6] and a large variety of neural networks [5, 8, 23]. Most of the behaviors that have been evolved on real robots have been fairly simple. Researchers have mainly tried to show the validity of their approaches by evolving behaviors such as obstacle avoidance, light seeking, and wall following. One of the important unanswered questions concerns the ....

.... of the problems with on line evolution mentioned above, a number of researchers working with real robots have tried to build simulators for their robots, and then evolved the control systems in simulation before transferring the evolved control systems to the real robots in the real environments [19, 23, 24, 12, 16, 32]. There have been di#erent approaches towards this. Jakobi et al. 16] from Sussex have suggested using a mathematical description of motor and sensor responses of the specific robot. Their simulator is based on a set of equations, that should model the real world physics, which are used by the ....

[Article contains additional citation context not shown here]

O. Miglino, H. H. Lund, and S. Nolfi. Evolving Mobile Robots in Simulated and Real Environments. Artificial Life, 2(4), 1996.

No context found.

Miglino, O., Lund, H., and Nolfi S. "Evolving Mobile Robots in Simulated and Real Environments." Technical Report, Institute of Psychology, C.N.R., Rome. (1995).

No context found.

Miglino, O., Lund, H., and Nolfi S. (1995). "Evolving Mobile Robots in Simulated and Real Environments." Technical Report, Institute of Psychology, C.N.R., Rome.

No context found.

Miglino O, Lund HH, Nolfi S (1995) Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434.

No context found.

Miglino, O., Lund, H., and Nolfi S.: Evolving Mobile Robots in Simulated and Real Environments. Technical Report, Institute of Psychology, C.N.R., Rome. (1995)

No context found.

O. Miglino, H. Hautop Lund, and S. Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 1996. To appear.

No context found.

Orazio Miglino, Henrik Hautop Lund, and Stefano Nolfi. Evolving mobile robots in simulated and real environments.

No context found.

O. Miglino, H.H. Lund .& S. Nolfi S., Evolving Mobile Robots in Simulated and Real Environments, Artificial Life, 2(4) (1996) 417-434

No context found.

Orazio Miglino, Henrik Lund, and Stefano Nolfi. Evolving mobile robots in simulated and real environments. Artifical Life, 2(4):417--434, 1995.

No context found.

O. Miglino, H. Lund, and S. Nol#, #Evolving mobile robots in simulated and real environments", in Arti#cial Life 2#4#, 1996.

No context found.

Orazio Miglino, Henrik Hautop Lund, and Stefano Nolfi. Evolving mobile robots in simulated and real environments. Artificial Life, 2(4):417--434, Summer 1995. ga95aMiglino.

First 50 documents  Next 50

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