Mel, B. W., Connectionist Robot Motion Planning: A Neurally-Inspired Approach to Visually-Guided Reaching. Perspectives in Artificial Intelligence, B. Chandrasekaran, Editor. Vol. 7. (1990) Boston: Academic Press, Inc.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....control has been to determine in what type of coordinates movement plans are generated in the brain. Teleological concerns argue strongly for planning in workspace coordinates, since the characteristics of a movement are most naturally specified in terms of the layout of objects in the workspace [17]. Based on this observation, in our method, instead of searching in joint configuration space as most existing motion planning methods do, we propose to run a direct best first search in the workspace, guided by a distanceto goal evaluation function. It incrementally computes an optimal path ....

....a crude measure of remaining distance to the goal at each step during planning that serves as a virtual potential field. The crucial point is that an easily computed potential field can be of good heuristic value, making it possible to search directly for paths without excessive backtracking [17]. While being inherently local, the algorithm does not suffer from local minima like most potential field based methods since the search is always continued with one of the six neighbors that is closest to the goal position. When no neighbors are accessible, the algorithm will backtrack. In ....

[Article contains additional citation context not shown here]

Mel, B. W., Connectionist Robot Motion Planning: A Neurally-Inspired Approach to Visually-Guided Reaching, Academic Press, Inc., 1990.

.... evidence that both sensor and motor regions are topologically organised and use population coded representations [6, 3, 2, 9] Representations are distributed over the activity of a whole population of neurons each of which respond over a range of inputs and have overlapping receptive fields (RFs) [9, 7] (figure 1(b) Such coding is efficient for generating coordinate transformations since it allows interpolation between nodes, and is robust to node failure and noise in individual neuron activations. Learning to form appropriate connections from the sensor to motor region is equivalent to ....

....defines the receptive field for the neuron. Bottom: The activation values of the population of neurons when representing an input value of 0.65. To learn the transformation between sensor and motor space requires training data covering the range of possible actions. Thus, most algorithms (e.g. [1, 9, 7]) go through a distinct training phase during which uniformly distributed random training data is generated and the inputs to the sensor region and outputs of the motor region are set to corresponding values from this training data (as if the data was generated by random motor actions) The ....

B. W. Mel. Connectionist Robot Motion Planning : A Neurally-Inspired Approach to Visually-Guided Reaching, volume 7 of Perspectives in Artificial Intelligence. Academic Press, 1990.

....is unable to cope with uncertainty. The environment has to be engineered to reduce the uncertainty to an acceptable level. The work to extend this paradigm to cope with uncertainty is still an active research area [3] Another approach is to have the robot system learn the task by itself. In [4], MURPHY, a robot motion planning system is presented. The system approaches the problem of visually guided reaching using connectionist method. MURPHY uses neural like units to learn the association between visual perception and robot arm motions in which it learns both forward kinematics and ....

B. W. Mel, Connectionist Robot Motion Planning: A Neurally-Inspired Approach to Visually-Guided Reaching. Academic Press, 1990.

....analog among computational learning techniques. Algorithms and data structures for nearest neighbor search enable the efficient retrieval of past instances (prototypes, or categories) according to their similarity to some new instance. Both nearest neighbor [5] and neural network approaches [12, 13] have proven useful for learning sensorimotor predictions from past experience. However, they differ vastly in their performance characteristics. Nearest neighbor learning converges immediately; it will converge to a final approximation once it has been exposed to all data points. Learning in ....

Mel, B. W. (1990). Connectionist robot motion planning: A neurally-inspired approach to visuallyguided reaching. Boston: Academic Press.

.... intelligence features like planification (geometric approach) Can87, CL90, FT87, Lat91, LTJ90] fuzzy and expert systems (rule based approach with structured information) HLG91, OMO 91, vdRvNLD90] neural nets (coding the knowledge into a black box with learning abilities) Kos92, Lee91, Mel90, NSA90, NW90a, NW90b, Wil93] Of course, conventional solutions have a definite advantage: there are theorems which state clearly when the solution exists and when it does not exist. But they are either implicit solutions or awkward solutions in the sense that they involve complex ....

.... is not available, new methods have been developed: planification (geometric approach) fuzzy and expert systems (rule based approach with structured information) neural nets (coding the knowledge into a black box with learning abilities) AM92, Can87, CL90, FT87, Lat91, Lau90, LTJ90, Lee91, Mel90, NSA90, NW90a, NW90b, vdRvNLD90] 4 Learning to control a system If we see the control of a system as a skill, we can distinguish two different parts: acquisition of the skill and refinement of the skill. These may be related to more classical notions such as system identification and control ....

B. W. Mel. Connectionist robot motion planning: a neurally-inspired approach to visuallyguided reaching. Academic Press, Perspectives in Artificial Intelligence: volume 7, 1990.

....training are well chosen and already encode the action to perform in such or such situation; in this application, the neural nets are a convenient way to control the mobile robots and allow some distributed control. 3.1. 3 Connectionist robot motion planning: Murphy This section describes Murphy [Mel90] a robot camera system associated with a connectionist architecture; the problem is to guide a multi link arm to visual targets in a cluttered workspace. No a priori model of the arm kinematics and of the characteristics of the vision system is assumed. Usually, when given a robot, one writes ....

B. W. Mel. Connectionist robot motion planning: a neurally-inspired approach to visually-guided reaching. Academic Press, Perspectives in artificial intelligence: volume 7, 1990. BIBLIOGRAPHY 49

....natural analog among computational learning techniques. Algorithms and data structures for nearest neighbor search enable the e cient retrieval of past instances (prototypes, or categories) according to their similarity to some new instance. Both nearest neighbor [5] and neural network approaches [12, 13] have proven useful for learning sensorimotor predictions from past experience. However, they di er vastly in their performance characteristics. Nearest neighbor learning converges immediately; it will converge to a nal approximation once it has been exposed to all data points. Learning in ....

Mel, B. W. (1990). Connectionist robot motion planning: A neurally-inspired approach to visuallyguided reaching. Boston: Academic Press.

....actual trajectory tended to oscillate about the desired trajectory. IV. DISCUSSION AND FUTURE WORK The utilization of optical flow simplifies the arm background segmentation problem significantly, assuming a static background. Unlike other approaches for manipulator control in the image space [Mel, 1991] which require identifying and tracking markers on arm joints such as LEDS, or grey level thresholding which is quite sensitive to ambient illumination, flow measures are much more flexible since they do not require explicit tracking, control of illumination or uniform backgrounds. The ....

B. Mel. Connectionist robot motion planning: A neurally inspired approach to visually guided reaching. Academic Press, San Diego, CA, 1991.

....during execution. Tan [20] employed a feature based sonar depth representation and a cost sensitive extension of ID 3 with 3 D objects. Bennett [2] worked in robotic grasping of polygonal 2 D puzzle piece task using explanation based learning and domain theories about uncertainty and grasping. Mel [10], Ritter [15] and Cooperstock have used Sigma Gamma Pi neural networks, self organizing feature maps and backpropagation, respectively, for learning visually guided control of robot arms for grasping. 2 Learning Algorithm First, we describe the learning algorithm, Density Adaptive ....

B. Mel. Connectionist robot motion planning: A neurally inspired approach to visually guided reaching. Academic Press, San Diego, CA, 1991.

....that are tangent to constraint surfaces in the configuration space [29] One possible solution to this limitation is to use vision based techniques to control motion in the remaining directions. Thus, much research attention has recently been focused on vision based control (see, for example, [2, 3, 9, 10, 15, 16, 22, 27, 30, 31, 32, 33, 34, 35, 37]) Although vision based control has been used successfully for a number of tasks (for example, in welding applications [1, 5, 21] none of the systems referenced above lend themselves to task level specification of goals, and therefore, there are currently no automatic planning systems that can ....

B. W. Mel. Connectionist robot motion planning: a neurally-inspired approach to visually-guided reaching. Academic Press, Boston, 1990.

....only works when the forward mapping is not redundant. If it is redundant, its inversion is not unique and the method employed by INFANT can not be used. Efficient learning of highly redundant mappings is discussed in Chapter 3. Examples of forward model learning are [Miller et al. 1987] and [Mel, 1990]. Miller is concerned with a robot that reaches for objects appearing in its visual field. Unlike Kuperstein, he learns a forward model first and then uses it to train a controller. Mel studies grasping for objects in the presence of obstacles. He uses the forward model to do a depth first search ....

B. Mel, Connectionist Robot Motion Planning: A Neurally Inspired Approach to Visually Guided Reaching, Academic Press, 1990.

.... 1993] or maintaining a heuristic confidence map [Thrun and M oller, 1992] Some researchers, in cases where the exploration is considered a secondary problem, provide the learner with a uniformly distributed training set, in effect assuming the problem allows unconstrained querying (e.g. Mel [1992]) An important limitation of the above work with dynamic constraints is that, for the most part, the methods are restricted to discrete state spaces. Continuous state and action spaces must be accommodated either through arbitrary discretization or through some form of on line partitioning ....

B. Mel. (1992) Connectionist robot motion planning: a neurally-inspired approach to visually-guided reaching.

....unit at this level can detect events such as the arm moving into: A) A particular range of the Y dimension. B) A specific region of the workspace (requires inputs from both the X and Y dimensions) C) Multiple regions of the workspace. A similar type of incidence coding scheme has been used by Mel [11], except that each sensory unit receives exactly one input from each dimension (e.g. X and Y) This results in only units of class B, which requires many more units to cover the entire space. The sensory units in the Planning layer receive input from the Target Object vectors. The dynamics of the ....

Mel, B. W., Connectionist Robot Motion Planning: A Neurally-Inspired Approach to Visually-Guided Reaching. Perspectives in Artificial Intelligence, B. Chandrasekaran, Editor. Vol. 7. (1990) Boston: Academic Press, Inc.

....since the practical consequence is that a tedious camera calibration phase is unnecessary. The utilization of optical flow simplifies the arm background segmentation problem significantly, assuming a static background. Unlike other approaches for manipulator control in the image space [12] which require identifying and tracking markers on arm joints such as LEDS, or grey level thresholding which is quite sensitive to ambient illumination, flow measures are much more flexible since they do not require explicit tracking. The next step in the capping experiment would be to incorporate ....

B. Mel. Connectionist robot motion planning: A neurally inspired approach to visually guided reaching. Academic Press, San Diego, CA, 1991.

....methods, local function approximation methods like Kohonen maps [Ritter et al. 1992] CMACs [Albus, 1975] Miller, 1989] radial basis functions [Poggio and Girosi, 1989] and back propagation neural networks [Rumelhart et al. 1986] store, retrieve, and interpolate between given data points. [Mel, 1990] combines a neural network approach with a depth first search of possible reaching strategies. Each method develops an efficient representation once a suitable set of input output pairs has been found. However, none of these addresses the problem of efficiently obtaining the data to be learned. ....

B. Mel, Connectionist Robot Motion Planning: A Neurally Inspired Approach to Visually Guided Reaching, Academic Press, 1990.

No context found.

Mel, B. W., Connectionist Robot Motion Planning: A Neurally-Inspired Approach to Visually-Guided Reaching. Perspectives in Artificial Intelligence, B. Chandrasekaran, Editor. Vol. 7. (1990) Boston: Academic Press, Inc.

No context found.

B. Mel, Connectionist Robot Motion Planning: A Neurally Inspired Approach to Visually Guided Reaching, Academic Press, 1991.

No context found.

B. Mel. Connectionist Robot Motion Planning: A Neurally Inspired Approach to Visually Guided Reaching. Academic Press, 1990.

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