A method is presented for calibrating the CyberGlove using a vision based system. Coloured dots are manually placed on the joints and the joint angles are then calculated by analysing the video. The gain and offset are then calculated using linear regression when there is a one to one relationship between the actual joint angle a and the CyberGlove reading, or a least squares method when there is a one to two relationship (that is, two CyberGlove readings can affect a single actual joint angle). The system provides good results when visually inspected.

An algorithm is presented for reconstructing the position of missing markers (for example when using the OptoTrak) based on knowledge of the distances between the markers.

A method is presented for calibrating the Cyberglove without requiring additional hardware. The subject holds the thumb and the index (or other) finger together, while moving the other joints. The hand is then approximated as a close kinematic chain, and the ratios of the bone lengths are assumed, and hence the only unknown joint angle is that between the thumb and finger. A least squares fit is used to minimize the calculated distance between the two fingers. The parameters in the optimization are the sensor gains, fixed offsets, the bone lengths and cross coupling terms (which allow for a joint angles to be dependent on the value of more than one sensor). Limits on parameter deviation are necessary to prevent trivial solutions.

A method is presented for finding the centre of rotation or axis of rotation for a joint from the positions of 3D markers (eg Optotrak). Unlike other methods, the relationship between the markers is not assumed - each marker only need maintain a constant distance from the centre/axis of rotation. The solution is a least squares solution and does not require setting any parameters.

A series of experiments were performed to test the accuracy of the CyberGlove. They found that noise had an insignificant effect on the results. They showed that using the standard calibration (and not calibrating for each user), reasonable differentiation could be made between a small number of divisions. For better differentiating capabilities, it is necessary to calibrate for an individual user - this can be achieved by measuring the data from the glove at known joint angles.

A method for calibrating an instrumented glove (the UTAH Dextrous Hand Master) is presented. The finger is modeled by an open-link kinematic chain. The end point of this kinematic chain is externally measured (using an Optotrak marker). Singular Value Decomposition is used to find the parameters. The poses used are carefully selected to use most of the joint ranges, and parameter scaling is used in the optimization procedure.

A model of the location of the major flesh points in the mouth is presented for speech production. The effect of coarticulation (where the position of a point during speech is affected by the previous or following phoneme) is estimated by measuring the amount of effort, defined to be the local curvature of the trajectory. The inclusion of coarticulation in their model produced better predictions.

The differences in playing on the piano two sequences that begin the same but have significant differences at some point were compared. It was found that in some cases there is anticipatory changes in the kinematics about one note before the divergence. As this was only sometimes the case, they concluded that such movements are executed in a strictly serial ordering as long as this is compatible with the task.

Coarticulation in ASL fingerspelling was studied. The joint angles of the hand were measured and discriminant analysis was used to classify postures and define a coarticulation measure. Substantial evidence was found of both assimilation (reducing distances between two shapes) and dissimilation (emphasizing the difference between shapes) - these effects were even found to apply concurrently in different joint angles. Assimilation was primarily found in the thumb and wrist joints, while dissimilation was found mainly in the PIP of the index and middle fingers. This discrimination may aid in recognition, as these joints have been found to be sufficient for 88% correct classification of letters. They suggest that the concurrent instances of assimilation and dissimilation argue against synergistic control, however, they reconcile this with the evidence for synergies by hypothesizing that there is a combination of a general tendency for coordination of all fingers with an ability for individuated control.

For recognizing an alphabet for fingerspelling, PCA is compared to using a subset of the joint angles. It was found that using a subset of joint angles was superior to using a similar sized PCA weighting vector in transmitting information about the pose. They suggest that hence synergies are not used as a primary control strategy for this task, and that recognition can be performed more easily and with less measured angles using this technique rather than PCA.

Anticipation in handwriting was studied by presenting to subjects the middle letter in a trigram and asking them to predict what the third letter will be (from two choices). Subjects could predict this with a reasonable accuracy (around 60%), however when the ratio between the radius of curvature and tangential velocity (ie the two thirds power law) is changed, the rate of accuracy drops significantly. They suggest that the ability to predict the next letter is based on an internal model of the movement. They also note that the prediction is based on the kinematic properties rather than the shape.

Neural networks are constructed for learning the joint angles of the index finger, based on the normalized number of EMG zero crossings, and a torque estimate based on the EMG (which is a combination of a number of EMG signals). Both methods modeled well the joint angles of the finger. Although considering the torque estimates rather than EMG was supposed to allow for different muscle cocontraction strategies, both methods performed statistically the same. It is hypothesized that the neural network based on EMG learned the cocontraction of parts of the movement.

This article presents an analysis of grasping by a multi-fingered hand and considers what ``intelligence'' is necessary for successful grasping. The intelligence is based on what prior knowledge is necessary of the object (e.g. only kinematic information, information about the mass, etc). It is suggested that successful grasping can be achieved with little intelligence by using a combination of learned signals combined with sensed physical values. For secure grasping, force feedback is not necessary, however for artificial grasping such feedback can be used instead of the learning process that humans may perform.

This study compared the effects of different tasks performed on an object with the performance of the reach and grasp movement before performing the task. Based on measuring movement parameters such as movement time, velocity and acceleration, they observed that the task to be performed does affect the prehension movement leading up to it.

A survey is made of the requirements of a machine hand for grasping, looking at human operability, manipulator dexterity and grasping robustness.

The influence of obstacles of grasping movement time (thumb / index finger) was considered for different models. It was concluded that their influence is best described by a model based on control of the thumb/index fingers, rather than a limitation of grip aperture.

An fMRI study was performed where subjects observed pictures of objects and had to decide in a forced choice task whether to use the object or grasp it. Use the object was divided into prehensile use (pinch or clench) or non-prehensile use (palm or poke), while the grasp condition was either pinch or clench. The left inferior frontal gyrus, posterior superior temporal gyrus and inferior parietal lobule (IPL) showed significantly greater activation in non-prehensile use compared to grasp. No areas were observed that showed greater activation for grasp. They suggest that this might be because computations for object grasping are a subset of the computation for using. A difference was only seen in the left IPL when comparing non-prehensile use and prehensile use. They conclude that the left IPL is important for storing knowledge of hand postures for functional object use.

In this work, a cybernetic hand, called the ``cyberglove'' is presented. The cyberglove has 6 actuators (motors), controlling the four fingers independently and the thumb. Each of the four fingers has three joints which are controlled by one ``tendon''. The thumb is controlled by two motors. The hand is able to perform opposition with the thumb, and can perform lateral pinch, cylindrical, spherical and tripod grasps. The high level control (i.e., selection of which grasp and amount of force) will eventually be based on EEG / EMG signals. The low level control is responsible for actuating the desired force. Some sensory feedback is also collected.

A set of visually computable grasp features was presented such as contact point arrangement and force equilibrium. They were used to build a neural network to predict whether a grasp will be stable. The training was performed by a robot which shook the objects to test their stability. This method does not require a model of the object to be grasped.

A measure is presented for task compatibility of a manipulator for certain task requirements (in terms of effecting or controlling velocity and force). The measure is based on the velocity and force ellipsoids. The transmission ratio of applied force or velocity in terms of joint coordinates to the same quantity in task coordinates in computed. The transmission ratios represent the amplification in force and velocity, while accuracy is represented by the reciprocals of these ratios. The compatibility index is based on summing the ratios or their reciprocal (depending on the task) in the appropriate direction.

A set of experiments were performed to test whether grasps are planned by generation or by recall. The experiment involved grasping and moving a cylinder to different heights. Initially, the postures selected were assumed to be such that at the end of the movement, the joints will be in mid-range (the end-state comfort effect). This means that the higher the position to which it will be moved, the lower it will be grasped. In movements that began where the previous movement ended to where the previous movement started, the initial posture was close to where they had grasped it previously at the end of the movement. From this finding, they suggest that movement plans are recalled as well as being generated.

This paper examined the relationship between the orientation and shape (different aspect ratios) of a cylinder, and the orientation and aperture of the grasping hand (consisting of precision grasps on the index finger and thumb). They found that the orientation of grasping was such that the cylinder was grasped (close to) along its principle axes, with more (68%) along the minor axis. It should be noted that grasping along one of the principal axes is the only stable grasp of a cylinder. They also found that the final hand orientation could be inferred after only 30% of the movement distance, whereas for the aperture this was only possible after 80% of the movement distance. They conclude that the findings confer with the theory that only the appropriate visual quantities are used in planning a movement, and that the errors observed are consistent with those due to the grasp being planned incorrectly due to a distorted perception of the cylinder's shape.

A review is presented of studies in human grasp choice, and analytical methods used for robotic grasping. Various categorizations and taxonomies used for human grasping are described, as well as expert systems. Algorithms for robotic grasp planning, including limitations of such models due to the assumptions made are presented. Different features that are optimized or used as constraints are described, and finally comments are made on the connections between the two.

For force production task, they define a mode, which is the forces produced by all the fingers as a result of voluntary force production in one finger. Multiple finger force production can be modeled by the superposition of modes but with a weight dependent on the the number of fingers used (to take into account force deficit). This model captures the behaviour of the enslaving effect for multiple fingers.

The ability to combine internal models for grasping was examined by measuing the peak grip force rate for lifting objects of the same appearance but different weight and their combination. In contrast to other studies, it was found that subjects could learn grip force scaling for two seperate weights simultaneously even when alternating between them. They suggest that this difference was because the objects were clearly distinct in the environment. In addition, they suggest the CNS may be able to additively combine two dynamic internal models to determine the necessary grip force for lifting the two objects together. It appeared that the subjects acted in a Bayesian way to deal with the uncertainty of the weights when they were combined.

A formalism is described for compliant motion, as an extension of Mason's hybrid control. It consists of selection of the task frame relative to the end effector, constraints on the force, velocity or tracking (detection of errors based on forces or velocities) in 6 dimensions in the task frame, additional task frame or end effector motion constraints, feedforward velocity constraints and task termination conditions.

A framework for implementing compliant robot motion is presented. The system receives as input the constraints as described in a previous work. It is based on a multidimensional position control loop embedded in a multidimensional force control loop.

A method of learning human grasps for telerobotics is presented. Four human grasps are recognized using magnetic trackers placed on four fingers. A Hidden Markov Model (HMM) is used for grasp recognition. The human posture is mapped to a (simulated) robotic posture using a trained artificial neural network.

The difference in grasping an object with different visual feedback was studied. The subjects were shown, before the movement, either the real object, a virtual computer rendered object or a symbolic presentation (using a mirror setup). The visual information was removed at the initiation of the movement. Haptic feedback was provided (using a real object). Different kinematic properties were compared (e.g. preshape aperture, grasp onset latency, movement velocity), and no significant difference was seen between grasping real and virtual objects (as opposed to pantomiming behaviour found in other studies).

A grasping experiment under the Ebbinghaus illusion showed that contrary to previous studies, the illusion affects grasping to the same extent as perception. It was shown that an alternate hypothesis of object avoidance cannot explain the results. They suggest that the same source is responsible for the illusion in both perception and in grasping. This reduces the evidence for a perception vs action hypothesis of brain organization.

When passively observed natural reaching and grasping movements, profiles of cortical excitability were in concordance with the kinematic profiles of the movements, and evoked greater corticospinal facilitation than the observation of unnatural movements. Depending on the type of perturbations, either no modulation was observed, or one similar to the natural movement. It is thus suggested that the resonant motor plan is loaded at the beginning and tends to complete itself regardless of changes in visual cues.

Three types of neural networks were compared for predicting the finger forces required in a torque stabilization experiment. The most effective one was a hierarchical two layer network, where first the virtual finger force was calculated, then in the second layer the finger forces were calculated. Input to the first layer was also available to the second layer. The performance was better than for a classical 3-layer network. They suggest that this supports the notion of hierarchical control of prehension.

The control of the fingers in grasping a sphere was studied under different conditions (varying the distance to the object and its size). Most of the time a tripod grasp was selected. This grasp consists of apeture components - the opening of the thumb and index/middle fingers which was coordinated, and a seperation component (between the thumb and index fingers) which was weakly coupled with the aperture component. They relate these findings to the use of the virtual fingers to form the grasp - one for the thumb and the other for the other finger(s).

The differences in grasping an object with the hand and with a tool (two mechanized fingers) were studied. Some kinematic features were preserved, while others were different. In particular, the same finger pre-shape was used for the grasp in both cases, but the temporal pattern of the movement was different (a pronounced velocity plateau, shorter opening phase and longer closure phase). Based on these results, they suggest that some grasp features are encoded independently of the effector used.

The increase in grip force observed in older adults may be due to the lower coefficient of friction of their skin rather than to compensate for greater instability.

The effect of displaying the word LARGE or SMALL on a block being grasped in a reach-to-grasp movement was studied. It was found that an effect was seen in the early stages of the movement, but the effect was seen less towards the conclusion of the movement. An explanation for this behaviour was that the meaning of the word affected the early planning stages of the movement, but do not affect the on-line control which uses different information.

Words representing large objects (such as apple) and small objects (such as grape) were presented to subjects before a grasping movement. Words representing large objects led to a larger grip aperture. The interference was apparent early in the movement and its effect diminished as the hand approached the target, which they explain as the result of on-line correction of the semantic effect. They consider this behaviour in terms of the distinction between motor planning and on-line control.

Precision grasp-like movements with the thumb and index finger were performed, and the joint angles, velocities and acceleration were measured with the CyberGlove. Inverse dynamics were then performed to estimate the joint torques, on which they performed PCA to joint torque synergies. Although the Principal Components for torque described less variance that those for joint angles, under different conditions (faster or slower velocity), the joint torques were observed to scale linearly with the velocity.

Muscle synergies during precision grip in humans was studied by looking at the EMG. It was found that rather than using a unique muscle synergestic muscle activation pattern for a particular task, the CNS appears to use flexible short-term muscle synergies. This variation does not explain the consistent and accurate behaviour observed in such grips.

A model for robot grasping is presented. Three different optimization criteria are suggested for producing high-quality grasps - minimizing muscle effort, minimizing the maximum applied finger forces (to prevent object damage), and maximizing the degree of uniformity between the fingers. By solving these optimization problems, suggested grips can be produced.

Three grasping quality measures are suggested - minimal muscle effort, minimum of the maximum applied finger forces, and minimizing an entropy-like function (which causes a uniform level of the contact forces). Subjects were asked to grip various objects, and the numerical values for these quality measures were calculated. These were compared with the subjects' perceived quality of the grip using a psychophysical magnitude estimation method. It was found that the measure of the uniform level of contact forces is dominant in the human quality sense.

This paper explores the paradox of separate channels for reaching and grasp formation and a holistic programming of such movements. To combine the two notions, it is suggested that the movements are organized on several levels. The individual channels are embedded into an internal model of the entire movement which exerts top-down control.

Grasp stability (of a multi-fingered hand) is examined by expressing it in terms of differential equations. The stability of the grasps is then determined by considering the Liapunov stability of the system of differential equations. Methods are then given for synthesizing stable grasps based on these concepts.

The coordination of hand movements and gaze was studied. Subjects fixated on on landmarks critical for control of the task, such as points where contact was made with the object. They did not fixate on the arm or the bar being grasped. They concluded that gaze supports the planning of the task by fixating on key points.

The control of grip stability was studied during pronation and supination movements of an object which has destabilizing torque dependent on the angle of rotation. It was found that the grip force for stabilizing the object increased directly with the destabilizing torque. As blocking sensory information from the fingertips did not significantly change the coordination, they concluded that feed-forward rather than feedback mechanisms are responsible for grip force control.

An algorithm is presented for learning to grasp using visual information based on a heuristic. Learning is used to improve the estimation of where to grasp and well as the measures of grasp quality.

The trajectories of the fingertips during grasping of 5 objects was studied. A good fit of the fingertip positions was found to a logarithmic spiral in the theta-r plane (and better than a polynomial in the x-y plane). The spiral was a good fit regardless of starting posture. More variance was seen for the thumb than the other fingers. Sometimes highly linear relationships were found between joint angles although not consistently. The lack of correlation found may be because the correlation is piece-wise rather than consistent over the movement.

The UCM approach is applied to a difficult multi-finger ramp force production task. The contributions of forces that contribute to the task force, and of moments in the frontal plane were considered as the hypotheses. The variance was partitioned into the component which does not affect the hypotheses (UCM) and the component that does. No difference was seen in the variance of the forces before learning, but a significant difference was seen in the variance after learning (i.e. less variance in the task component). The variance in the moment stabilization became worse after learning (this is an unavoidable consequence of better force stabilization).

The properties of the grasp stiffness matrix are examined. It is shown that a stiffness matrix is conservative if the matrix is symmetric and satisfies a certain differential condition. In general a conservative stiffness matrix is Cartesian space will be nonconservative when transformed into joint space using a configuration dependent Jacobian (and vice versa).

Three issues involving multifingered hands were examined. A method is presented for selecting internal grasp forces to produce a stable grasp. It is based on specifying suitable constraints (e.g. friction, joint torque limits) and finding the configuration that is furthest from violating any of these constraints. Also presented is a method for finding motion of the fingertips (e.g. rolling) due to movement of the object. Finally, a method is presented for finding the workspace of a hand/object pair, that is, the range of manipulators for a particular configuration of contact points on the object and locations of the contact points on the fingertips.

A non-dimensionalized composite grasp index was constructed, based on a stability index, a grasp uncertainty index, a maximum force transmission ratio index, a task isotropy index, and a stiffness mapping-based isotropy index. Each index was appropriately normalized, and has a weighting factor based on the relative importance given to this component. By altering the weighting factors for each index in a simulation, different optimal grasps were produced.

In order to determine the optimal grasp, a series of performance indices were defined. These indices are a stability grasp index (how close the grasp points are to a regular polygon), an uncertainty grasp index (how far away the grasp points are from edges), a maximum force transmission ratio index (based on the force ellipsoid and the desired force direction), a task isotropy index (distance from singularities) and a stiffness mapping-based grasp isotropy index (based on the grasp stiffness). These measures are normalized (by dividing them by the difference between the maximum and minimum possible values) and thus also non-dimensional. Different weights can be given to the different indices depending on the task.

A method is described for planning the necessary stiffness for various grasping and manipulation tasks. The stiffness of the grasped object is related to the stiffness of the joints through the grasp matrix. The desired stiffness geometry for the task in object coordinates can then be transformed to determine the necessary joint stiffness and/or geometry of the hand. Various examples are given.

A pneumatically controlled glove is described that can be used for rehabilitation. The five fingers of the glove, which is worn by the subject, can be extended together by the bladder which is sewn onto the palm side of the glove. The pressure of the bladder is controlled by a servo valve connected to a computer. Its use was demonstrated on a stroke survivor in a virtual reality simulation where the patient has to grasp objects, assisted by the glove.

The effect of semantic category (living vs non-living objects) and size on a reach-to-grasp task was examined. Inconsistent results were found regarding the difference in speed between living and non-living objects, but the size was found to have a significant effect on the kinematic parameters. The effects of distractors was also noted.

The adaption to a novel, velocity dependent force perturbation was found to be different depending on the specified goal. When subjects were asked to perform a spatial goal (continue to the target), their movements became curved but returned to reach the final point. In constrast, when subjects were asked to maintain the same effort, the deviation increased throughout the movement, resulting in large endpoint deviations. A significant after effect was only seen with the spatial goal.

Synergies have been observed for finger force production, that is, that other fingers will compensate for an error or variation in the force produced by one finger. This studied asked the question of whether the other fingers can correct for timing errors, i.e. if there are timing synergies. Evidence was not found for such synergies, rather, if one finger sped up, the others were also likely to speed up.

An algorithm is presented for learning hand gestures using a Hidden Markov Models (HMMs). Twenty joint angles from the hand are used as input. They are first preprocessed by dividing them into gestures, resampling, applying a FFT and creating a single vector from the data. This is used as the input to the HMMs - there is one for each of an alphabet of gestures, and the one with the highest probability is selected if the classification is strong enough).

PCA is used to find a lower dimensional representation of static grasp postures using 18 joint angles of the fingers. They also consider a non-linear PCA, which allows non-linear mappings between the principal components and the desired posture. This produced slightly better results than the regular PCA, but is a more complex procedure.

A neural network was constructed that could predict the effects of force production in multi-fingered force production tasks. Unlike optimization techniques, this model accounts for force deficit and enslaving observed experimentally.

Several quality measures are defined for multi-fingered grasps. They present quality measures based on the grasp matrix, G. They introduce general quality measures, based on the smallest singular value of G, and the volume in wrench space. They also define a task-oriented quality measure, based on the task ellipsoid (force ellipsoid). The specification of the task ellipsoid for a task is based on experience with the task and similar tasks.

A frame invariant measure is defined for compliance grasps, and an interpretation of the stiffness matrix is given.

A method is presenting for calculating the stiffness matrix using the Hertz model. They contrast this to the linear spring compliance model that is commonly used but is not supported by experiments, and the coefficients must be determined experimentally.

To examine what is learnt during manipulation of unstable objects, an experiment was performed where the subjects had to balance a simulated inverted pendulum. When the arm posture was changed, the results were better when the effects of arm torque were matched to the first condition. From this result, they suggest that the subjects learnt the necessary joint torques rather than a general model of forces. A further experiment found that the advantage of training was object specific, based on comparing two different tasks with similar forces but different visual cues and requirements.

The relationship between the joint angles in the arm were studied during a reaching and grasping task of an object at different orientations. A linear relationship was observed between upper arm torsion and the torsion of the forearm relative to the upper arm. They conclude that a combination of upper arm, forearm and fingers are used to specify the orientation rather than by using separate transport and hand orientation components.

The difference in some kinematic parameters of the hand during different tasks was examined. Significant differences were seen, mainly in the relative time of the peak velocity of the wrist. The tasks that required greater precision has a longer deceleration phase. Based on these findings, they suggest that movement planning be relatively specific to the task.

A technique is described for modeling the PIJ joint angle and the forces applied at the finger tip (normal and shear forces but not moments) based on the patterns of blood volume beneath the fingernail. The blood volume is measured using LEDs and photo detectors. Shear forces are measured to an accuracy of 0.5N root mean square (rms) error, normal forces with 1N rms error and PIJ angles with 10 degrees rms error.

In the first part of this book, Mason analyses different types of contacts, using the notation of screws, twists and wrenches. He uses this to define which hand grips are stable. The grip transform is introduced as a way of transforming forces applied by the fingers to the force applied to the object. Stiffness control as a way of controlling the hand is also presented, as well as the design of a robotic hand (the Stanford/JPL hand). The second part of the book by Salisbury looks at the mechanics of grasping and pushing.

A system for combining several modes of communication for instructing a robot grasping task is presented. Speech and vision are combined to specify the command. In the manipulation stage, the robot moves by switching between different arm and hand modes (eg. approach, shape, grasp, release).

A simulation of grasping was presented, based on stored postures. Separate postures are used for the arm and the hand. The constraints, in order, were to avoid collisions, spatial accuracy and movement cost reduction. A search is performed through the postures and a goal posture is selected, and a via posture if necessary to avoid a collision. The predictions of this model were compared with experimental results. It was predicted and found experimentally that larger object sizes correspond to smaller aperture overshoots. A further prediction that larger objects cause the moment of maximum aperture to occur earlier was not seen experimentally. This model is limited in the sense that it is only a kinematic model but does manage to capture many of the properties of such movements.

Objects are modeled by primitives (boxes, sphere, cylinders and cones), and based on these primitives, one of two grasp preshapes is selected. As well as the type of preshape, the location and orientation of the wrist (6 parameters) and its orientation (2 parameters) are specified. A number (50 to 100) of possible preshapes are generated based on some simple rules defined by the type of object. These preshapes are tested by moving the hand to the object and closing the hand until contact occurs. The grasp is evaluated using a stability measure. Infeasible grasps (for example, because of an obstacle) are discarded. The best grasp can then be selected.

A simple rule is presented for predicting the relative durations of the opening and closing phases of the hand during prehension. They propose the duration of each phase is proportional to its amplitude (do and dc), i.e. To/Tc = do/dc. The relative time (To and Tc) to maximum aperture is thus determined by the ratio of opening and closing apertures. An experiment showed that 96% of timing variance is account for by this rule.

A Parameterized Self-Organizing Map neural network is used to learn the 20 joint angles of a hand only based on the locations of only the fingertips. This allows approximate reconstruction of the joint angles of the hand from only a small amount of information (the locations of the finger tips).

The role of the thumb in force production tasks in different grasp configurations was considered. When the thumb acts in parallel to the other fingers, it acted similarly to the other fingers (in that the force applied was less than if it applied force by itself). However, when it acted in opposition to the other fingers, the peak force was much larger than when it applied force by itself. They conclude that in some configurations (i.e., in parallel to the other fingers), the CNS treats the thumb as a fifth finger with respect to force deficit and enslaving, although the muscles used for the thumb do not have the relationships that exists between the other fingers. From this, they suggest that the magnitude of interaction between the fingers has a significant neural and not only biomechanical component.

A model for how infants may learn to grasp is presented. The model consists of several modules specialized for the task (a virtual finger layer, a hand position layer and a wrist rotation layer). The selected grasp is determined based on an input (the location of the target) according to a probability distribution. The feedback to the learning is based on a reward signal based on the grasp stability. The model successfully ``learns'' to grasp, similar to that of infants. Based on the model, they suggest that infants can acquire grasping rather than innately possessing it and that initially grasping is an open-loop process.

The coupling of the arm and finger movements during prehension tasks was tested by looking at the kinematics from an experiment with a double-step paradigm involving a task where the subject had to reach a grasp a dowel. The velocity profile of the wrist was bell-shaped, with 2 peaks seen in the perturbed trials. The aperture of the grip was seen to increase to a maximum, then decreased to close on the dowel. In perturbed cases, often two peaks were seen in the aperture profile. It was noted that the each peak aperture followed a wrist velocity peak, however they concluded from a statistical analysis that the two components are not systematically coordinated but rather time-coupled in some way.

An algorithm was devised for efficiently planning stable grasps (for a Barret hand) on undeformed superellipsoids. The set of possible grasps was parameterized using four parameters - two for the starting position of the palm, one for the roll and one for the spread of the fingers. For each superquadric, 3,600 grasps were generated that span the space. SVM regression was used to efficiently compute the grasp quality, and this quality was maximized for given shape parameters. The algorithm succeeded in producing stable grasps, however it was for a simple hand and class of shapes.

Grasp planning is considered here as the solution to an optimization process with certain constraints. The optimization condition is to minimize the difference between the finger force magnitudes. The conditions (such as force and moment equilibrium, contact forces) applied, are dependent on the type of grasping (normal, with controlled sliding, or regrasping).

Reach-grasp-lift movements were performed on virtual objects of identical size but different simulated mass and coefficient of friction with the floor. Haptic feedback was provided with a Phantom haptic feedback device. When the object was more stable (greater mass, or higher coefficient of friction, the contact force was greater. They suggest that this means the stability of the object is learned, and is used in planning the movements. Movement to more stable objects also showed different kinematic properties, in the form of larger hand apertures and velocities. If no haptic feedback is provided, the movements are similar to those of unstable objects, hence for movements involving stable objects, haptic feedback is needed to avoid unrealistic movement features.

The difference in grasping a real and a virtual ellipsoid was studied. Grasping the virtual object showed greater deceleration and variability - this is probably due to the lack of contact at the end of the the motion. Furthermore, the type of grasp selected was dependent on the amount of visual information given.

A mathematical analysis of force production in pushing a pivoting stick was performed to determine what is required to maintain static stability. The hand rotational and translation stiffness can be used to stabilize the stick. It is suggested that such a strategy is generally used by humans for force-production task. Such analysis can also be useful in tool design.

The properties of neurons in the dorsal premotor area F2 of macaque monkeys was studied during grasping tasks. The neurons were classified as purely motor, visually modulated or visuomotor, depending on whether they were affected object presentation, motor action or both. Some neurons showed preference for the type of grasp (e.g. side grip vs precision) and others for the size of the object.

The effect of changing the centre of mass and handedness was compared on the forces applied by the fingers during grasping tasks. For each finger the normal and tangential grip forces were measured. Similar patterns of forces were used despite the unpredictability of the centre of mass and with different hands. It was also noted that the normal forces exerted by the fingers are synchronized and usually in-phase or out-of-phase. This suggests that some sort of synergies are used for coordinating the fingers during grasping tasks.

A model of motion planning is presented based on stored postures. An initial goal posture is selected that satisfies certain constraints (eg it is close to the object, it doesn't collide with the object, travel costs are low), and postures close to this initially selected posture are also generated. The best of these is selected, and if collision will occur a via posture is also generated. The posture of the hand is first generated, then that for the arm. The movement is then executed. The predicted movements predicted well several features of such movements. It should be noted that the model is for movements in a plane, although they suggested how to extend it to a 3D model.

This paper formulates a description of dynamic manipulability, analogous to manipulability (velocity) ellipsoids, which gives the relationship between joint velocity and end effector acceleration. The effect of velocity is taken into account in the formulation.

Hand motion during reach to grasp of real, virtual and remembered targets were studied. By using PCA, it was found that two principal components can account for >75% of the variation. The first PC is made up of the extension and flexing of the fingers. The second PCA, which begin about 70% of the way into the movement accounted for the extension of the digits.

A series of experiments were performed looking at the accuracy of adjusting finger span to various objects. Different permutations were made - to size, shape, distance, orientation and finger configuration. None of these factors had a major effect on the accuracy, contrary to the findings of other studies. Whole hand movements to grasp a cube were also measured using the CyberGlove. Almost all the variance in these movements could be described using two principal components - the first remained fairly constant throughout the movement, and the second represented the bending of the fingers that varied throughout the movement. The small number of principal components needed to describe the movements however may be due to the specific task (grasping cubes).

It was found that the joint angles representing the posture of the hand while gripping imagined targets did not vary independently between objects. Rather, most of the variance could be described using a much smaller number of components. They suggested that this means that the hand posture is controlled with a few postural synergies.

It was found than when gripping concave and convex objects, the hand gradually mould to the shape. The posture of the hand discriminated between the shapes well before contact, although the discrimination was incomplete at the time of peak aperture. It is suggested that this is because this parameter is not fully specified until later in the movement.

The effect of object shape and visual feedback during grasping was studied by an experiment where subjects had to reach and grasp objects in different visual feedback conditions. They suggest from the results that at least two motor processes occur in grasping. The first is a preshaping of the hand (about 45% of the movement time), and the second is a slower grasp modulation to refine the grip to its final shape. Movement duration increaded with lack of visual feedback.

A comprehensive review is made of grasp synthesis algorithms for robotic grasping. Grasp properties are categorized according to grasp dexterity, equilibrium, stability and dynamic behaviour. Algorithms are suggested for synthesizing grasps with the desired properties.

This paper claims that their grasping model based on constraints on the end effector is just as effective as more complex models based on the posture of the arm and hand in predicting the main features of grasping movements. From this they say that postural constraints are not important in trajectory formation of reach to grasp movements.

A model of grasping is presented that rather than modeling the movement as two separate parts (transport and grip) models the movement on the entire movement of the thumb and fingers. This is based on the notion that it is the thumb and not the wrist that is transported during grasping. The movements are then planned using minimum jerk trajectories but with the assumption that the fingers and thumb approach the object perpendicularly. This model predicts several features observed in prehension movements, such as that the object size affects the maximum aperture but not the movement of the wrist.

Finger movements during typing were studied, using principal component analysis on each joint separately. This showed that only a few (2 to 4) principal components were needed to explain most of the variability of each finger. Cluster analysis was also used to test hierarchical relationships, and showed evidence of patterns between the joints, or synergies.

The number of synergies involved in some hand manipulation tasks is considered using Principal Component Analysis, based on the assumption that the first few principal components describe the main synergies involved in a task. They found that 6.5 Principal components are necessary to describe most of the variance for different manipulation tasks. For a task involving moving all the joints individually, they found that only 8.5 principal components are needed (due perhaps to biomechanical coupling). These results are higher than in simpler grasping studies, but do not show that the neural controller eliminates many of the synergies it has access to. Furthermore, different synergies were observed for different tasks and between subjects. Based on these results, they suggest a task-optimal control strategy (optimizing only parts of the movement related to the performance) gives a better explanation that simplifying the control.

A system is described for robot imitation of grasping movements. The system tracks the hands and fingers using a stereo camera. The tracking is performed based on Gabor jets, which measures the similarity of an image fragment to a template. The grasping is based on tracking the location of the index finger and thumb, and is implemented using a gripper.

A model was presented for planning reach-to-grasp movements. Three components were planned - hand position, grasp aperture and hand orientation. Each component is planned based on a difference vector between the current and desired position. Coordination between the components is achieved through a common (increasing) gating signal which ensures that the components end approximately simultaneously. An additional feature is introduced to the aperture control, called self-inhibition, which accounts for the tendency of the hand to return to a relaxed position. This model accounts for a nuber of features observed experimentally for such movements. Perturbations of the object are handled by altering the common gating signal. This model is implemented as a neural network.

The dextrous capabilities of the raccoon hand are presented. Although the raccoon does not have a truly opposable thumb, it is capable of dextrous manipulation. It achieves this by avoiding fingertip grasps, and instead using the palm or more commonly some other fixed surface (such as the ground) to constrain the object. The scissor grasp, grasping using abduction between the fingers is also sometimes used to constrain objects. They also grasp and regrasp an object a few times before being up, in this way it is believed that the dynamics of the object is learned and it is placed in a convenient orientation and location. They also make use of the two hands for grasping and manipulation. It is suggested that such techniques could be used in robotic hands which are less dextrous than the human hand.

A review is presented of the issues involved in multi-fingered grasping. He reviews the technicques involved in grasp stability, finger force distribution, and grasp compliance.

This paper studied the temporal organization during the activity of drinking from a bottle with a glass using two hands. It was suggested that the movement is organized such that synchronization will occur at critical times during the movement. This would be part of a top-down control mechanism for motor execution. Additionally they found that the temporal structure was common across different modalities (different forms of pantomime and with the real objects).

This review suggests that motor plans are represented in higher coordinate structures which then coordinate the necessary interactions at the lower executional levels. The context of the motor tasks influences the particular organization used (for example, compliant and unrestrained movements show different curvature).

Task-dependent preshapes are used as a way of simplifying robot grasp planning. A preshape is selected (by the user) depending on the task (e.g. precision, lateral or manipulation). The preshapes have parameters - these are fit such that the aperture is proportional to the expected grasp distance, but kept as small as possible. The finger movements are then generated to close on the object, using a proximal or distal strategy. This method de-emphasizes stability analysis, rather assumes that the selected strategies will lead to stable grasps.

A Hidden Markov Model is presented as a way of recognizing and emulating human gestures. This allows the invariants patterns in the movement to be found.

A system was developed for using a CyberGlove along with force sensors to describe the posture and force to apply for telerobotics. It was found that the grip size was primarily controlled by changes in the MCP angle, while the main force exertion is from the thumb and index fingers. It was suggested that these primary parameters can be used to specify the robot grasp.

A comparison was made of reach to grasp movements between normal subjects and subjects with Cerebellar damage. Those with the Cerebellar damage performed worse for isolated reach, and grasp, movements, although they did not worsen the parameters of these movements when they were combined, although other deficits were seen - more variation in the movements, and a separation between the reach and grasp components. This decomposition strategy is believed to be a default strategy for these subjects. It was concluded that the cerebellum is probably involved in the control of combined reach and grasp movements.

This paper explains a system of classifying manipulative hand movements (coordinated movements of the fingers to manipulate an object). The joints are assumed to be coordinated such that they are in-phase or anti-phase, and so can be expressed in terms of another joint. The movements are segmented on-line by detecting the ``folds'', and the 16-dimensional vector representing the joint angles with respect to the most active joint is classified using an ART network. High rates of recognition are achieved.

Manipulative hand movements are assumed to be coordinated, and hence straight lines are expected in phase plane (when two joint angles are plotted against each other). These straight lines are detected using the Hough transform and used as a basis for segmenting the movement.

The forces applied by the fingers in a task where the subject had to keep a handle vertical under differing load and torque conditions were studied. The moment required to keep the handle vertical was provided about 50% by normal forces and 50% by shear forces. The index and little finger torques were found to depend mainly on the torque, while the middle fingers depending on both the applied torque and the load. Additionally, antagonist movements were always seen, even when they are not mechanically necessary.

A Neural network model was used to explain the forces applied by the fingers in tasks requiring application of torque and force. Optimization of the finger forces could not explain the results seen, due to the effect of ``enslaving effects'', where a finger that is not required to produce a force is activated because of commands given to a different finger.

The synergies involved in a force and torque production task were studied. They defined a synergy as conjoint changes of finger forces and moments during multi-finger prehension tasks. Evidence was observed for use of synergies as a way of resolving the redundancy. For example, the adaptations made were of the synergy as a whole, rather than as a minor change.

It is claimed that, as is used in robotic control, humans perform superposition when performing prehension. They looked at grasping a handle with a prismatic grip with different applied torques. They observed no correlation between the forces needed to prevent the object from slipping and for maintaining the object orientation and hence concluded that they are controlled by separate commands. They also found that the finger force changes associated with the changing of one of the parameters did not depend on the other factor, and hence concluded that the two commands can be summed.

The synergies involved in a precision grip were reviewed. It was claimed that there are two independent commands, one to prevent slipping and one to maintain the rotational equilibrium, and superposition can be used to combine these commands. Due to the large space of forces that can be applied, these synergies can only identify a subspace of solutions, and some other mechanism needs to make fine adjustments to meet the task requirements.

The connectivity between two bodies is the number of independent parameters needed to describe the relative locations of the two bodies. It can be calculated from the mobility and redundancy of the system. The connectivity is used for classifying into three types of grasps - power grasps, constrained motion grasps and free motion grasps. These classifications can be used in grasp synthesis.

A set of canonical grasps are specified in terms of the location of the fingers and the object, and a graph is defined that specifies the possible transition between these canonical grasps. A manipulative movement is then programmed by finding a suitable path in the graph from the starting grasp to the final grasp. This method was tested experimentally using a three fingered robotic hand.

A quantitative measure of multi-fingered grasps is presented. It measures the capability of the grasp to hold the object under external disturbances. The measure is calculated from the set of contact wrenches. It allows grasp analysis and synthesis.

A method is presented for recognizing manipulative hand movements as part of a larger system for programming a robot by demonstration. The movements are segmented based on force sensors (to determine contact with edges). The grasps are further segmented between static and dynamic grasps. Dynamic grasps are classified according the taxonomy of Elliot and Connolly. Classification is performed based on joint angles measured with a data glove and a Support Vector Machine (SVM) classifier is used. High classification rates (around 90%) were achieved.

It was shown that sensimotor adaption where the scene was rotated 60 degrees is learnt between certain movement categories, namely tracking and pointing. The magnitude was larger from pointing to tracking. They suggest that adaptation is located in the brain before the divergence for different movement categories.

A simple sensory feedback scheme that operates in task space is described for controlling arm movements. This technique avoids the need to perform inverse dynamics or deal with excess degrees of freedom. The damping terms in the feedback equation can be selected to prevent self-motion and to cause the velocity profile to be roughly symmetrical and bell shaped. This model with appropriate selected parameters is simulated for a 4-joint arm model making movements in a horizontal plane.

Double step movements in grasping are investigated, where the target (a disc) sometimes changes size after the ISI time following the initial target presentation. In particular, they consider the aperture of the grip (the distance between the thumb and index finger). They consider whether the change in trajectory is due to cancellation, superposition or amendment. They conclude that neither of the three is a good description.

A biomechanical model of the index finger is presented that can predict the tendon extensions and forces based on the trajectories and applied forces. It is based on the combination of models of tendon extensions and forces. The unknown parameters are solved using a recursive Newton-Euler method under the additional constraint of minimizing muscle stress to solve the otherwise under-constrained problem.

The kinematics of point to points movements of the index finger moving in a plane were studied. It was found that the movements were not straight, and the path was dependent on the direction (i.e., a to b has a different path to b to a). From this they suggest that the trajectory plan must not be solely kinematic and must take into account mechanical properties. However, they did not control for starting or ending posture, which could also be the cause of different paths in the two directions.

A simple gesture recognition system is described based on the coordinated movements of landmarks

A model is presented for the control of a non-rigid object by the arm. As opposed to studies involving adaptation to perceptual or mechanical perturbations which are parametric perturbations, this is an example of a structural perturbation because new equations are required to describe this system rather than changing the parameters of an existing dynamic equation. The movement is of a virtual mass attached to the hand by a (virtual) spring. The boundary conditions of the movement are on the initial and final position, and that the hand and object velocity and acceleration should be zero at the start and end of the movement. This provides 10 independent boundary conditions (because of dynamic coupling), and a 9th order trajectory is used to describe the motion of the object and a different 9th order trajectory describes the hand motion. A 9th order polynomial is predicted by minimizing the mean squared crackle (5th derivative of position) of the object trajectory. For fast movements, this model predicts hand velocity profiles with 2 peaks, a direct contradiction of the predictions of the minimum jerk model. Such velocity profiles were experimentally observed. They note that this model, which they call the optimally smooth transport principle, is a descriptive rather than explanatory account of the movement. They suggest from anecdotal evidence that visual feedback is required to learn the task, because unlike movements involving the limbs, no proprioceptive feedback is available.

An experiment was performed with virtual targets, with single step and double step movements, where the target could change in size or position 300ms after the target appeared. The change in object size effected the kinematics of the grasp but not the transport component, while a change in object position changed the kinematics of both the grasp and transport components. The correction time was found to be distinctly different for the grasp and transport components. It was also noted that in cases of double-perturbation (change of position and size), these responses can not be thought of as a combination of two single-perturbed responses. They conclude that their data is consistent with a model for prehension based on two mutually coupled channels (for grasp and transport).

This paper classifies hand movements in terms of the types of synergies (simultaneous or sequential), as well as the patterns and digit groupings and use.

A route planning system is described that uses online control to determine the current state without an explicit world model or path plan. The route is planned using a dynamic model in terms of the angular acceleration. The goal acts like an attractor, and the obstacles like a repeller. Multiple objects can be simulated by linear combination. The routes predicted by the model were similar to those performed by humans in a Virtual Reality experiment.

A gyroscope was used to alter the dynamics of a hand movement. It was found that the hand path did not change, but that the configuration of the arm was altered. As the subjects learned the movements, the arm gradually returned to its normal configuration, implying that different forces were generated. The normalized peak of the kinetic energy did not increase with the learning - from this they suggested that kinematics and kinetics might be mutually optimized.

This set of experiments considered the effect of object affordances on grasp selection. An object affordance is a motor representation that causes particular types of interaction, such as the size of the section to be grasped or the object's weight. One theory states that only the relevant affordance for the task influences the grasp selection, while a second theory claims that the grasp selection will be influenced by all the affordances of the object. If this second theory is true, then the grasping of objects should be affected by object affordances which are not part of the current grasp. This was observed in a series of experiments, and based on this, the author suggests that objects has a single motor representation which is used in grasp planning and implementation.

A neurophysiologically plausible model is proposed for movement recognition. The model is also quantitative, allowing its predictions to be tested. It is based on two pathways, for form and for motion, analogous to the ventral and dorsal streams. Each pathway is a hierarchical model that begins with low level details - local orientation detectors for the form pathway and local motion detectors for the motion pathway. These low level details are combined hierarchically to give representations at different levels, until recognition can be performed. These levels are related to different brain areas. The model is capable of explaining the results of many existing studies.

A study of independence of finger movements found that when asked to move one finger, motion in the other fingers was also produced. This lack of individuation was the same for dominant and non-dominant hands, and less independence was seen when the frequency of cyclic movements was higher (for 3Hz compared to 2Hz). The unrequested motion may be due to passive mechanical connections, the organization of multi-tendonded finger muscles and from neural control.

Joint motion was measured with an ultrasound based motion analysis system. It was found that there is a linear relation between the proximal and distal interphalangeal joints, equal for flexion and extension. The ratio is 1 (PIP) to 0.76 (DIJ).

The relationship between muscle strength and noise was examined for different muscles in the arm during a torque matching experiment. The force was measured using a force transducer. The relationship between muscle strength and muscle noise for each muscle was calculated based on the maximum voluntary torque production. This was compared with the results of a muscle simulation, where the output of the muscles was the summed result of muscle twitches caused by a spike train with a Gaussian interspike interval distribution. The number of motor units and the spike train noise were varied. It was observed that as joint strength increases, the coefficient of variation decreases exponentially. The simulations were able to accurately model the data, from which they conclude that stronger muscles with more motor units have a lower coefficient of variation.

A review is made of strategies for dealing kinematic redundancy (or as he describes it, kinematic abundance). He considers methods based on relationships between the variables (such as using PCA) and those based on some form of minimization.

Four models were compared for direct and via-point pointing movements - minimum work, minimum angular jerk, minimum travel cost, and Donders' law. In terms of absolute error, Donders' law gave the best description of the data.

Different types of forward models are presented and it is noted that is necessary to first define what type of internal model (in terms of input space, output space and its structure) before evidence for or against the existence of such models can be considered.

In a study of adaptation to varying force fields during reaching movements, it was found that subjects were unable to adapt to a time-varying force field while they were able to adapt to a velocity-varying field. They speculate that the system that adapts movements to external forces cannot use a temporal representation.

A planning scheme for a 3 DOF robot finger is presented, based on the human finger. The key features is that the distal interphalangeal (DIP) joint and the proximal interphalangeal joint (PIP) are linearly related. It is compared to planning the movement in order to maximize a manipulability measure. They conclude that the requirement of interphalangeal coordination can produce natural trajectories.

The development of arm kinematics of infants is studied. Through the first two years, the path become nearly straight and the number of "movement units" decreases, and the movements become unimodal. Still, there are considerable differences between the movements of a 3 year old and an adult.

A series of experiments were performed to support the theory that the central nervous system uses probabilistic models during sensimotor learning. Using a virtual reality setup, subjects were believed to have learned the distribution of lateral shift which had a Gaussian distribution. This was tested by using different feedback conditions. The trajectories observed support such a model over a model where subjects estimate the average lateral shift, as well as a model where they learn a mapping from the partial feedback to an estimate of the shift. Subjects were also capable of learning more complicated distributions, such as a mixture of two Gaussians.

The strategies in a multiple finger force production task were compared between normal subjects and subjects with Down syndrome (DS). It was found that a simpler, sub-optimal strategy was used for controlling the force applied with the DS subjects, where they did not compensate for errors between the fingers. However, practice had a considerable effect on improving finger coordination with such tasks.

A finger force production task was examined. As seen previously, the variance in forces related to the task was much lower than the variance in forces unrelated to the task. Similar results were found between a discrete task (ramp force production) and an oscillation task. From this they concluded the synergy organization is the same between such tasks. It was also noted that the stabilization of force was only possible within a certain range of values for the force. It was suggested that this may be because error correction of the forces involves time delays that are too long to achieve stabilization.

This paper presents the Uncontrolled Manifold (UCM) hypothesis for analyzing variability in motor control. This hypothesis assumes that there is a subspace, the ``uncontrolled manifold'', for which the variance is not controlled, but only along ``essential'' directions that do not belong to the UCM. Hence high variability can be shown as long as it remains inthe UCM.

The extent of motion enslaving between the fingers is observed. Finger movements were restricted so that only the distal interphalangeal joints could move. Considerable enslaving was observed - the motion of one finger caused the slightly delayed movement of one or two slave fingers, with amplitudes for some fingers greater than 60% of their peak amplitude. The index finger was the most independent. Several possible explanations are given for this phenomenon.

The predictions of different types of control schemes in a pole balancing task were studied. This task was chosen because memorized motor commands cannot be used, rather, closed-loop visual feedback is needed. During some trials with a virtual pole, the visual feedback was blanked-out, but the subjects succeeded in maintaining to balance the pole as well as when they had feedback (up to a certain amount of time). From this, they suggest that there is a forward model in the control loop. They show that a delay uncompensated control model and a Smith predictor model can be eliminated as feasible control hypotheses.

A robot learning algorithm is developed based on spline fitting using via points. In an example using a simple game, via points are extracted from human demonstrations, in terms of Cartesian coordinates, and joint angles (from a 7 DOF arm). In reproducing the trajectory for the robot manipulator, the Cartesian coordinates must be exactly reproduced while the joint angles of the robot should be as close as possible to those of the human. The trajectory for the robot is generating by fitting splines to the via points.

Learning during adaption to a destabilizing perturbation of a knob turning was examined. After time, the subjects learned to move accurately under the perturbation and their overall kinematics and performance measures returned to close to what they were before the perturbation. They suggest that their observations support the hypothesis that subjects adapt by learning to make more accurate primary movements (i.e. without corrections) and the other measures (such as smoothness) can be explained as a secondary effect.

A method is presented for reconstructing the joint locations from markers placed in arbitrary locations on the limbs. The method is based on calculating the transformations from one limb to the next and finding a point that remains still between the transformations. A best-fit solution is found to take into account the noise in the system.

Subjects were able to learn two force fields that were applied in opposite directions during a centre out task. The force fields were velocity dependent, and one operated in a clockwise direction, the other counterclockwise. As opposed to other studies whether the opposing fields were presented in an alternating sequence, or in blocks, and subjects did not succeed in learning both models, they showed that using random presentation the subjects were able to adapt to both fields. The force fields were accompanied at the start of the movement by audio and visual cues to indicate which field would be presented.

A model for how infants may learn to grasp is presented. The model consists of several modules specialized for the task (a virtual finger layer, a hand position layer and a wrist rotation layer). The selected grasp is determined based on an input (the location of the target) according to a probability distribution. The feedback to the learning is based on a reward signal based on the grasp stability. The model successfully ``learns'' to grasp, similar to that of infants. Based on the model, they suggest that infants can acquire grasping rather than innately possessing it and that initially grasping is an open-loop process.

The performance during typing was measured when the index finger was anesthetized and in a control situation. The average kinematics remained the same - they suggest this is because typing is executed by an open-loop system. However, typing errors increased sevenfold, and much more variability was observed. Regression analysis showed that the endpoint variability was mostly predicted by variability of the starting position. They suggest that the starting position was poorly predicted due to the lack of tactile feedback (such feedback aids in accurately measuring the finger posture).

The movements of a 7 DOF human arm are recorded and used to teach a neural network about reaching movements. General movements can then reproduced by a virtual mannequin that have properties of human movements without them being explicitly specified.

The role of variability in motor behaviour as more than simply noise is examined. They describe how the strategy of first separating movements into deterministic and random components can lead to missing out important features of the movement. They suggest that the variability may even be more revealing than the invariants of the motions, and suggest tools for analyzing the variance.

A 3D model of the human index finger is presented that can be used for estimating the muscular forces involved in free finger movements.

A biomechanical model of the four fingers in the hand is described. Each finger is considered as an open chain of rigid bodies (the bones) connected at the joints. The movement of these chains is controlled by the muscles (of which 25 are considered) through the tendons. This model was used to predict the maximum voluntary grasping force for different sized cylinders.

A neural network is used to model the movements of a 3 DOF finger. The redundancy is dolved by assuming the PIP and DIP angles are equal. A minimum jerk velocity profile is assumed to produce smooth movements

An automatic method is presented for detecting biological movement from Johansson displays, based on maximizing the joint probability function of the position and velocity of body parts.

This paper compares the minimum crackle model for object motion with a minimum jerk hand constraint. The minimum crackle model, presented elsewhere, requires that the mean-squared-crackle of the object be minimized. However, they show that if the object is rigidly connected to the hand, then this does not reduce to the minimum hand jerk as would be expected. Also, if multi-mass objects (objects connected by springs) are considered for manipulation, then minimum crackle would be insufficient and higher order derivatives of position are required. Instead, they minimize the hand jerk, under the dynamic constraints of holding the object. Their predictions for multi-mass objects are much better at predicting such movements than an extension of the minimum crackle model. They note that this model is only presented for 1D (i.e., it does not predict the path), and does not require taking into account the inertial properties of the arm.

A review is made of using optimization principles in motor control, contrasting open loop and closed loop models. Open loop optimization models are presented as representing the average behaviour, while closed loop models allow integration of sensory information, i.e. optimal feedback controllers.

The constrained minimum-jerk model is presented in this paper. This model is based on the minimum-jerk model, but the jerk cost is minimized when the path is specified (rather than just the end or via points). They found experimentally in a range of movements that this model better predicts the velocity profile than the different versions of the 2/3 power law. They suggest that this model is applied over a small sliding windows (of approximately 1 second) rather than globally.

A model for planning movements is presented based on independently planning the geometrical and temporal components. Under different speed conditions, the the position-orientation hand paths were found to be similar. The kinematics on the movements were found to be dependent on the initial and final postures.

A gender classifier for point light display of a human walking on a treadmill is constructed. PCA is performed on the 3D locations of 15 markers and the first 4 Principal Components describe 98% of the variance. Each movement is then described as a trigonometric function in terms of the mean posture, the first 4 eigenpostures (principal components), the fundamental frequency and phase shifts of the eigenpostures. The eigenpostures were similar across subjects. PCA is performed again on these descriptions, and a linear classifier used on this representation. 90% correct classification is achieved (compared to 76% by human observe