Design of Customized Rehabilitation Aids

Design of Customized Rehabilitation Aids Vijay Kumar and Ruzena Bajcsy

GRASP Laboratory, University of Pennsylvania, Philadelphia, PA (USA) email: kumar/[email protected]

William Harwin

Applied Science and Engineering Laboratories, A. I. duPont Institute University of Delaware, Wilmington, DE (USA) email: [email protected]

Abstract We present the component technologies that are essential for rapidly designing and producing customized rehabilitation aids for people with motor disabilities. We show that methods traditionally used in robotics and computer vision can be used to formulate and solve many of the problems that are encountered in automating this process.

1 Introduction There have been many attempts to design robot systems [2, 5, 10, 19, 20] for aiding people with physical disabilities in their activities of daily living. These include robotic aids that operate in such structured environments as oces and act as vocational assistants [20], as well as systems which are designed for operating in unstructured environments [19]. Systems that belong to the rst category allow disabled people to pursue a vocation and to attain a certain level of autonomy [32]. The robotic manipulation tasks encountered here are considerably simpler because the environment is structured. In contrast, systems in the second category include such general purpose aids as a six degree-of-freedom robot manipulator mounted on a wheelchair with dierent end eectors [19] and a wheelchair with two arms that can be used for manipulation as well as for climbing over or stepping across obstacles [34] and are potentially superior for functioning in unstructured environments. However, these versatile aids appear to be be less acceptable to people with disabilities. There are many reasons for the lack of success of general purpose robotic aids in this community [28]. Such systems tend to be very complex and

therefore unreliable. They are also prohibitively expensive. Robot arms move slowly and awkwardly, but at the same time appear to move too quickly in tasks involving contacts with the environment or the user. People with disabilities are often leery of electromechanical and possibly computer controlled aids that rely on sophisticated technology. Finally, one of the main obstacles is the diculty that the users have in controlling such complex systems [21]. In contrast to a prosthetic limb where the user is in intimate contact with the limb and therefore has proprioceptive feedback (Doubler and Childress [6] have called this extended proprioceptive feedback), the user typically only has visual feedback from a robot system. While haptic interfaces is an active area of research, there appear to be inherent limitations with the technology that preclude simple and cost-eective mechanisms for force and touch sensing [4]. Consider, as an example, the task of feeding for a person who has little or no control over the upper extremities. The person may use intact musculature in the legs or in the head and neck, either directly [10, 27] or by generating myoelectric signals [3], to control a robotic arm to feed with utensils. However, the robot arm is an expensive product and is dicult to use. In contrast, consider the solution shown in Figure 1, called the Magpie [7], in which an articulated passive feeding mechanism allows the user to manipulate a feeding utensil via a set of cables. It is reminiscent of the early teleoperators which had no actuators, sensors or computers [33]. The Magpie is an example of a telethesis, a device that is closely coupled to the user and acts as an \extension" of the person. This telethesis has four degrees of freedom and is controlled through movements of the leg and foot. Although this device is human

Figure 1: The articulated mechanism for feeding in a foot-controlled feeding device (Magpie) designed at the Nueld Orthopeadic Center in Oxford [7]. powered, it is possible to use actuators to assist the human so that it remains passive and therefore stable [14]. Finally such devices are simple, reliable and inexpensive. Clinical trials with the Magpie [7] show that there is a high degree of consumer acceptance. In this paper we consider the customized design and manufacture of a class of rehabilitation aids that is typi ed by the telethesis shown in Figure 1. The design and manufacture of such products presents a novel problem. Because each person presents a unique neuro-physiological picture, there is considerable variation of performance and function and therefore, it is essential to design tools that are speci c to that person. It is necessary to involve the customer in any design process, but this is especially true for rehabilitation aids. Furthermore there are biological changes that occur over time, and it is necessary to allow for adjustments and maintenance or to rapidly redesign and manufacture a new product. Traditional models for product development and manufacturing focus on low-cost, high-volume products. In contrast to this, the manufacture of rehabilitation aids requires the infrastructure and

technology to design and produce a wide array of quality products each of which targets speci c market needs. While agile manufacturing [29] makes it possible for a designer to move quickly from a preliminary design concept to a prototype, it does not speci cally address the need to customize products to individuals. The primary focus of this paper is on the component technologies that are essential for rapidly designing and prototyping such customized products. Regardless of the speci c product class, the rst important step in the production of a customized product is the quantitative assessment of the needs of the individual. This involves the acquisition of geometric, kinematic, dynamic and physiological information about the individual which is necessary for developing design speci cations and for detailed design. Because the product volume for customized products is likely to be small, the manufacturing cost must be kept low. Thus, there is a need to automate the process of measuring the customer and designing the product from speci cations derived from these measurements. In addition, there is always pressure to provide the product quickly and be able to respond to the consumers' needs rapidly. We consider as a test product a telethesis that could be used by a quadriplegic for feeding, and possibly for typing on a keyboard, for turning pages of a book and for shaving. More than 10,000 spinal cord injuries (SCIs) result in quadriplegia or paraplegia each year in the U.S. alone. Of the population of 300,000 people in the U.S. with SCI, 55% are quadriplegics, and 58% are between 16 and 30 years old [22]. There is often a high level of motivation for disabled people to learn to eat independently [28]. Thus there is a signi cant population that will bene t from an inexpensive, customized feeding aid. While most of the discussion in this paper is limited to rehabilitation aids, there are many other examples of customized products. All products that are worn on our bodies, or more generally, products that depend on prolonged contact with human bodies for their functionality, need customization at some level. This class of products includes wrist braces, computer interfaces (keyboards, joysticks), eyeglasses, helmets and sports equipment. The level of customization that is justi able for each product depends on the cost, time for development and for production, the production volume and ultimately pro tability. A discussion of successful product development is beyond the scope of this paper.

Instead, our main objective is to focus on engineering technologies that are required to support the rapid design and production of customized goods and services tailored to the needs of a speci c consumer.

3. Rapid Design and Prototyping: to take a preliminary design, convert it into a detailed design and quickly produce prototypes for evaluation and later for production.

In the rest of this paper we describe each of these steps in greater detail, with particular reference to the computer integration. The end target is The design process for rehabilitation products to design and prototype a successful customized that are customized to a person will involve a product. number of steps, as shown in Figure 2. Of these, We use the example of a proposed headthere are three stages that are of central interest controlled feeding device to discuss the basic ishere: sues in customized design and prototyping. A head-controlled feeding mechanism is particularly useful for people with quadriplegia. When quadriplegia is the result of a spinal cord injury, the upper extremities and even the torso may be impaired, depending on where the lesion occurs. However, the quadriplegic generally has good control over his/her head and neck movements. In contrast to the foot and ankle control for the telethesis shown in Figure 1, we consider a head-control interface, as shown in the schematic in Figure 3, with which head and neck movements can control the extension and retraction of the cables and therefore the movement of the feeding utensil.

2 The Design Process

DATA ACQUISITION

IMAGING, FORCE/MOTION MEASUREMENT

MODELING, QUANT. ASSESSMENT

DIAGNOSIS

PRELIMINARY DESIGN SPECIFICATIONS

PRODUCT EVALUATION

SIMULATION

DETAILED DESIGN

CONSUMER FEEDBACK PRODUCT EVALUATION VIRTUAL PROTOTYPING

A 2

RAPID PROTOTYPING

PRODUCT EVALUATION

Figure 2: Design and rapid prototyping of oneof-a-kind rehabilitation products.

AA AA 1

1. Data Acquisition: to measure and observe the user while performing the desired tasks in the appropriate environment and to develop quantitative models for assessment by a therapist or a designer and for providing design speci cations; 2. Virtual Prototyping: to simulate the user, the product, and their (physical) interaction through powerful workstations with highresolution graphics throughout the dierent stages of product design; and

Figure 3: A schematic of a head control interface. The user can move his or her head, neck and chin to control multiple (two are shown in the gure) degrees of freedom.

3 Data Acquisition It is necessary to measure the capabilities and needs of the individual, his/her environment and to describe the task in quantitative terms in order to generate the speci cations for the design problem. For example, the custom design of a headcontrolled telethesis for feeding requires the measurement of the geometry of the head, the kinematics of the head and neck, and the forces that the person can apply with his/her head. Similar measurements may also be required for the feeding task (for example, the ranges of motion of the spoon or fork and the forces that are encountered during the task). For customized design we require, in addition to geometric measurements (shape, size), information about the kinematics Figure 4: Geometric information of a human and dynamics of the individual. Each of these head obtained by a Cyberware 3030PS scanner. categories is discussed next. The information is obtained by scanning the head from dierent angles and registering the dierent images. The information of the chin is used to 3.1 Geometric measurements make the customized prototype in Figure 5. The typical input will be range data from a given human part (shoulder, arm, wrist, neck, head) obtained through the use of two monocular cam- there is a need for methods for representing geoeras and/or from range nding scanners. Figure 4 metric models at various levels of granularity deillustrates the kind of geometric information that pending on the desired tolerance and the comcan be obtained from a commerciallyavailable op- plexity of the part [16, 17]. tical range scanner. In addition to modeling the human body part, such measurements can also 3.2 Kinematic measurements provide information for the design of products that are worn by the user. For example, a helmet The basic needs are the accurate measurement or a rigid head band worn on the head, or (as of the three-dimensional motion of moving body shown in Figure 5) a customized chin strap can parts (for example, the head and neck) withbe automatically prototyped with such geomet- out interfering with the dynamics of the movric information. Geometric models for such com- ing bodies and to develop methods for modeling ponents can be created automatically, and from the three dimensional movements. The available these models, CAD/CAM models can be easily measurement systems include optical (accurate to generated [17]. These models can be used either 1-2 mm) and electromagnetic (accurate to 0.3for prototyping (and manufacturing), as shown in 0.8 mm) sensors [23]. Another possibility is a Figure 5, or for inspection and quality control. mechanical manipulandum (an articulated, lowThe main challenges in obtaining the required inertia linkage with high-resolution position senmeasurements turn out to be important research sors, one end of which is attached to or moved by problems in computer vision. First, several range the individual) which is accurate to within 0.01 images must be taken of the object from dierent mm [8]. vantage points and merged into a single surface Regardless of the measurement device, it is description. It is necessary to automatically de- necessary to automatically determine kinematic termine where to position the range camera next, models for articulated human limbs with minimaking sure all of the range images are perfectly mal sensing (see, for example, [23]). This probaligned, and merging the range images into a sin- lem is similar to the kinematic calibration probgle surface model [25]. Second, robust techniques lem encountered in robotics [12]. While the are needed for concurrently segmenting and t- medical community has developed sophisticated ting geometric models to each segment from data techniques for measuring joint movements, many obtained from multiple viewpoints [16]. Finally, of these techniques are time consuming and re-

Figure 5: A customized chin strap designed for the subject shown in Figure 4: (a) a ProEngineer solid model; (b) a stereolithography prototype. quire clinically trained professionals and expensive equipment. Also, in this community, the kinematic models for limb movements are generally limited to ranges of motion at dierent joints. In contrast, we need fast, non-invasive measurement systems and methods for developing analytical models of limb movements. And frequently, we are not interested in the anatomy of internal features (such as the cervical vertebrae in head/neck movements), but we are only interested in capturing the kinematics of the gross motion. Our approach to automatically obtaining kinematic models is based on methods developed in a doctoral dissertation [23]. To obtain a kinematic model of, for example, head and neck movements, an electromagnetic sensor is strapped to the user's head. The user is asked to move his/her head while measurements of the head position and orientation relative to a xed base are recorded. If the kinematic chain connecting the head to the torso has less than six degrees of freedom (less than three degrees of freedom for planar movements), the six (three for the planar case) independent measurements obtained from the sensor must satisfy appropriate closure equations. These consistency equations allow us to infer the structure of the kinematic chain. The cervical spine is incredibly dicult to model because it consists of segments whose relative motions are roll-slide motions (and not pure rotations or translations). However, it turns out that the complex motions of the head and neck are adequately modeled by a kinematic serial chain consisting of only revolute joints. A model for head and neck movements in the sagittal plane is shown by the planar serial-chain

linkage in Figure 6 (a). The movement of the reference frame attached to the sensor (and therefore to the head) with respect to a xed inertial frame is characterized by a p  1 vector of constant parameters, x, and a n  1 vector of joint variables, q. In this case, x consists of the link lengths, a0 , a1, and a2 , and the angles, 0 and 3 . The variable q consists of the joint variables 1 and 2 that parameterize the two-parameter motion. For any position of the head (and the sensor) we can write three (six for spatial models) closure equations of the form: f (x; q) = 0; i

i 2 1; 2; 3

The model is completely described by the vector x and therefore we want to estimate x. Since q varies with the position of the head, each measurement gives us three equations with n additional unknowns. Thus m measurements yield 3m nonlinear equations in p + nm unknowns. In this case, p = 5 and n = 2. If m > 5, a nonlinear least squares algorithm can be used to determine the best- t 2 ? R linkage. A typical example is shown in Figure 6 (b). While the solution is not unique (the optimization problem is not convex), it is not dicult to obtain very good models. In the example shown in the gure, model matches the experimental data to within 0.01 inches and 0.2 degrees. Similarly, for general, spatial motions it is possible to estimate best- t n-parameter models where n < 6. For more complicated spatial models, the numerical methods work better if we can eliminate the variables q from the closure equations (which one can always do if n < 6). It turns out that a simple variation of the inverse

x SENSOR COORDINATE SYSTEM

θ3

y a2

θ2

AA AA AAA (a) a1

y

θ1

θ0

a0

x

(b)

(c)

Figure 6: Head movement in the sagittal plane: (a) a simple 2? model; (b) typical experimental data and the best- t model; and (c) the movements recorded over multiple trials plotted in joint space. (Figures from [23].) kinematics procedure for general serial chains [26] allows us to do this. Once the constants x are obtained, the variation of the n joint angles (q) can be obtained by inverse kinematics for each of the measurements. This procedure is quite general and can be used to construct models for a wide range of nparameter motions, where n must be less than 6 for three-dimensional movement and less than 3 for planar motions. Frequently, n turns out to be very small for voluntary movements. For example, if we look at the variation of the joint variables in our model with the head movement by plotting the motion in joint space (Figure 6 (c)), we can see that the motions appear to follow the same curve in joint space and further, this curve is closely approximated by a straight line. Our experiments with healthy as well as disabled people (subjects with spinal cord lesions in the cervical spine) indicate that head movements in the sagittal plane are actually 1-parameter motions [23]. Another approach which is not limited by the number of degrees of freedom involves the analysis of the deforming silhouette of moving limbs in a controlled studio-like environment [13]. In this work, no a priori model of the human body is assumed. Nor is it necessary to perform the dicult task to segment the image of the human body into limbs. The underlying model is a composed deformable model that allows for gross three-dimensional movements as well as local deformations. The basic procedure involves hypoth-

esizing a suitable underlying anatomical structure and then to test this hypothesis against the observed images. The technique applies to humans of any anthropometric dimension. An important direction for future research is the development of algorithms that will perform the data reduction and provide analytical models that can be used for designing the appropriate products.

3.3 Dynamic performance To measure the performance of an individual, it is often necessary to measure the kinematics of three-dimensional motion while measuring the forces and moments that can be exerted at the end-segment. The motion and force measurements, when appropriately transformed to joint coordinates, re ect joint motion ranges and strength characteristics. There are very few general-purpose instruments or general techniques to obtain such information. Our approach has been to use a manipulandum [8, 11] or a robot [27] to measure forces and moments that can be exerted by a hand during manipulation. A similar approach has been used to measure forces and moments that can be applied using head and neck movements by quadriplegics and healthy subjects [30]. This measurement combined with a suitable kinematic model for head movement yields a complete model of the static strength characteristics of the individual. Once such a model is available, the design of the feeding linkage can be formulated

Figure 7: A virtual prototype of a feeding device and the design interface. fairly precisely. A natural extension of this work is the dynamic modeling of the head/neck movements where one could obtain impedance models for a range of head positions. However, this is not essential for the application considered here. Other major goals for future research are a general framework for modeling and representing information on human performance, and algorithms for data reduction and representation.

4 Virtual Prototyping Virtual prototyping is the process of design, analysis, simulation and testing of a product within the computer, and using the results to re ne the concept and redesign the product before making a physical prototype. Over the last decade, high speed computer graphics workstations have proven to be very eective in allowing visualization of three-dimensional complex systems [1]. With advances in robotics technology, the potential for developing haptic interfaces that allow the user to feel forces exerted by the virtual environment (in addition to seeing the environment) has been successfully demonstrated [9]. As comput-

ers become faster and as more sophisticated actuators and sensors are developed, computer interfaces will enable the user to feel, touch and see the virtual product in a virtual environment. For customized design and prototyping, it is essential to integrate virtual prototyping with data acquisition. With the measurement of the user, the task and the environment, we can create accurate dynamic models (speci c to the user, the task and the environment) and investigate the virtual creation and installation of a customized virtual product on a virtual human user as an integral part of the engineering process. Consider again the example of a feeding device. To evaluate candidate designs, it is useful to create a simulation of the user and the mechanical system as shown in Figure 7. The mechanism that links the human head to the feeding device is not shown in the gure. The designer can experiment with dierent kinematic coupling mechanisms and see how the movements of the user are translated into the movement of the end eector or the spoon. Three-dimensional graphics provides visual information about the design, while a real-time dynamics simulation package elicits information about the forces and

the velocities that are required of the human head and neck to eectively accomplish feeding. By linking to an appropriate physiological database one can verify the feasibility of the required head and neck motions and also investigate possible sources of discomfort or trauma with the virtual prototype before clinical tests are performed. Being able to develop a virtual prototype of the product also allows the consumer to use and evaluate the virtual product in an appropriate virtual environment before the designer commits to the expense of creating the physical prototype. In the rehabilitation engineering domain, the designer may miss important constraints due to a lack of anity with the consumer [28]. As shown in Figure 2, consumer feedback (and evaluation by experts such as therapists) during the virtual prototyping phase and the redesign of the product in response to this feedback at a very early stage can ensure the success of the product and possibly avoid building multiple physical prototypes and incurring the resulting expenses. To this end, in addition to allowing consumers to test the product by testing the virtual prototype, it may be bene cial to provide haptic feedback so that a user can feel the dynamics of the product. Current work is directed to developing a head mounted haptic interface [27] (see Figure 8) in which the user wears a helmet that is attached to a six degree-of-freedom manipulator. For example, the manipulator can then be controlled to simulate the dynamics of the feeding device. But the current state of the art with haptic interfaces [4, 9] shows that much needs to be done before such devices can made safe, reliable and eective.

5 Rapid Design and Prototyping The design process can be divided into a concept development and system-level design phase and a detail design phase [31]. By rapid design we mainly refer to speeding up of the detail design phase, which includes the speci cation of the geometry, materials, and the manufacturing process for each component. A key issue during the detail design phase is to synthesize the coupling mechanism that will take a feasible and comfortable voluntary movement of the user and transform it into the motion required to perform a speci ed task. We note that selecting the voluntary \input" movement may be a dicult process and involves physiological

Figure 8: A head-controlled haptic interface for a virtual environment. The user's helmet is attached to the PerForce, a six degree of freedom manipulator (Cybernet Systems, Ann Arbor, MI). The manipulator can be controlled to simulate the dynamics of a virtual prototype. considerations. The process of \designing" the task is similar to the problem of task planning in robotics and is by no means trivial. However, once the desired input and output motions are identi ed, the basic problem can be formulated as shown in Figure 9. The telethesis can be decomposed into an input subsystem that is coupled to the user and driven by his/her motions, and an eector subsystem that performs the desired task but is driven by the input subsystem through an appropriate transmission. We outline the general procedure and some important considerations for designing the telethesis when the output motion is a 1-parameter (but possibly complicated and three-dimensional) motion. The input motion can be mathematically modeled as a multi-parameter motion. In the kinematic model for head/neck movements discussed earlier, the model was a serial chain with axial joints and the parameters corresponded to joint variables. While the input motion is, in general, a complicated three-dimensional motion, in many cases, this motion can be described by a simpler 1-parameter motion. This was seen to be true for head/neck movements (Figure 6 (c)). The input motion is denoted by u and the parameter by  in Figure 9. The parameter  cannot be arbitrarily chosen. It should be possible to design a link-

HUMAN INPUT MOTION, u

DATA ACQUISITION, MODELING

INPUT SUBSYSTEM

λ

u(θ1, θ2,...,θn)

MODEL REDUCTION

TRANSFORMATION σ (λ)

σ

u(λ)

EFFECTOR SUBSYSTEM φ1(σ), φ2(σ),...,φn(σ)

DESIRED OUTPUT MOTION y(φ1, φ2,...,φn)

Figure 9: The synthesis of a one degree-of-freedom telethesis for a desired input and output motion. age for the input subsystem that will \extract" the parameter . Physically, this means the input linkage should be such that one of its joint displacements must be proportional to  so that it is possible to drive the eector subsystem by coupling it to this joint. The design of the eector linkage is also an interesting problem. While a six-degree of freedom serial chain linkage, such as those seen in robots, will suce, it is clearly not required. Given the set of tasks that need to be performed, it is necessary to synthesize a linkage that will execute those tasks. By restricting ourselves to serialchain linkages we can formulate the design problem in the manner of [15]. By allowing for cabled transmissions in serial chain linkages, we obtain a wider class of solutions for the eector subsystem. There are two important design considerations. First, it is bene cial to keep the number of joints in the system to a minimum. Second, it should be possible to drive the eector subsystem with a single joint input, . Given the input and output (eector) subsystems, the only remaining task is to design the transformation between  and . This transformation can be designed to provide appropriate mechanical advantage and velocity ampli cation along the trajectory, y(). A possible eector subsystem for a feeding mechanism is shown in Figure 10 as an example. The linkage can be used to scoop up the food from the plate with any approach angle and bring the food to the user's mouth as he/she pitches his/her head forward. The mechanism has three degrees of freedom driven by cables labeled 1, 2 and 3 in the gure. 1 ; 2, and 3 are the corresponding cable extensions. The nominal yaw movement of the head, 1 , causes the linkage to rotate about a vertical axis and translate in a horizontal plane so that the spoon is always in the line of the sight of the user. The nominal pitch movement of the head, 2 , drives a planar 2 ? R open chain whose joints are coupled linearly so that the spoon performs a planar motion that involves scooping up

Figure 10: A candidate feeding mechanism that can be mounted to the back of the wheelchair. the food and bringing it to the mouth. The nominal roll movement, 3, causes the spoon to pitch about a transverse axis. The three independent parameters 1; 2; 3 are coupled to three input parameters 1; 2 ; 3 through a linear transformation (not shown in the gure). While most of the discussion here has been limited to kinematic synthesis, it is clear that there will be many applications in which the task forces will be signi cant and will impact the design problem. Also, we may not be able to decompose the multi-input, multi-output design problem into several single-input, single-output design subproblems. Even if this decomposition is

Cyberware range scanner

Video cameras

GEOMETRIC MODELING

KINEMATIC MODELING

3-D model

Surface mesh

Manipulandum

DESIGN INTERFACE

Motion

simulation

DYNAMIC ANALYSIS

VISUALIZATION (JACK, GeomView, ProEnginer)

DESIGN OPTIMIZATION

CAD/CAM ProEngineer

Off-the-shelf components

Customized components

RAPID PROTOTYPING

Figure 11: The user interface for the designer. possible, much of the literature on linkage design is limited to kinematic synthesis of closed chain linkages and is of very little use here. The synthesis of serial-chain mechanisms with cabled transmissions for specialized tasks, and in particular, the synthesis of multiple degrees of freedom mechanisms, is an important direction for future work. For rapid design and especially for customized products, it is essential to adopt a computer integrated approach where the designer can access and manipulate various heterogeneous pieces of information pertaining to geometry, kinematics and dynamics. At the heart of our design package is a graphical user interface (see Figure 11) which also acts as a server to support the interactive design and analysis processes. The front end is developed using Jack, a package for human body simulation [1]. The key idea is to have a generic request procedure which enables any of the component design/analysis packages or modules to call another package to obtain relevant information. Thus, information from any data acquisition, virtual prototyping or simulation module can be easily displayed on the visualization package. Finally, since the modules operate on dierent machines/architectures, ecient communication protocols between separate

processes (relying on Unix TCP/IP calls) are employed. This graphical server allows a modular approach to software development and enables the designer to interact with each module at different levels. Currently, the modules consist of the two data acquisition modules for geometry [16] and kinematics [13, 23], and the virtual prototyping module described earlier. The user interface allows the designer to create a personalized simulation of a particular individual with the relevant geometric and kinematic characteristics (see Figure 7). The designer can prescribe a desired trajectory (positions and orientations) for the simulated human agent using a mouse or a joystick. It is possible to see the simulated human execute the motions while conforming to the kinematic, dynamic and physiological constraints that are characteristic of the individual and while being subject to the dynamics of the environment (the feeding linkage, in this example). Our discussion of the design of rehabilitation aids was primarily focussed on the problem of synthesizing customized linkages that take as an input human motion and generate appropriate movement required for speci c tasks. Engineering design requires an integrated approach combining such diverse methods as reliability analy-

sis, models for costing [31], and optimization [24]. A detailed discussion of these issues is beyond the scope of this paper. Further, we did not attempt to discuss rapid prototyping, the process of quickly making a physical prototype of a product from a design in order to evaluate or test the product. We refer the reader to [18] for some of the challenges and research problems in this area.

6 Concluding Remarks

We presented the essential components necessary for the computer-integrated manufacture of lowvolume customized products such as rehabilitation aids for people with motor disabilities. We also brie y described some of the open research problems in the areas of data acquisition, virtual and physical prototyping, and systems integration. Many of these problems are not very dierent from problems encountered in robotics, computer vision and databases. The basic concept of customized production and the computerintegrated approach to design and prototyping is applicable to a wide range of products from clothing, sports and recreational equipment to computer workstations and machine interfaces.

Wheelchair Mounted Object Manipulator for Quadriplegics, Proc. 4th Int'l. Conf. on Rehabilitation Robotics, Wilmington, DE, June 14-17, pp. 103-106. [4] J.E. Colgate and J.M. Brown (1994), Factors aecting the Z-Width of a Haptic Display, Proc. IEEE Int'l. Conf. on Robotics and Automation, San Diego, CA, May 8-13, pp. 3205-3210. [5] J.L. Dallaway and R.D. Jackson (1993), The RAID Workstation for Oce Environments, Proc. RESNA Int'l. '93, pp. 504-506. [6] J.A. Doubler and D.S. Childress (1984), An Analysis of Extended Physiological Proprioception as a Prothesis Control Technique, Jrnl. of Rehabilitation Research and Development, Vol. 21, No. 1, pp. 5-18.

[7] M. Evans (1991), Magpie: It's development and evaluation, Technical report, Nueld Orthopeadic Center, Headington, Oxford, England OX3 7LD, 1991. [8] G. Garvin, M. Zefran, E. Henis and V. Kumar (1994), A study on optimality criteria for two hand reaching tasks,

Proc. 13th Southern Biomedical Engineering Acknowledgments: The support of National Conf., Washington, DC, April 16-17. Science Foundation, grant number MIP 94-20397, is gratefully acknowledged. Thanks are due to Mohamed Ouerfelli for his work on the modeling [9] V. Hayward (1995), Performance Measures for Haptic Interfaces, Proc. 7th Int'l. Symof head and neck movements reported in this paposium on Robotics Research, Hirshing, Gerper. We also thank Venkat Krovi, Rich Pito, and many, October 21-25. Ioannis Kakadiaris for their help with the illustrations. We are grateful to Jean-Marc Vezien and [10] J.R. Hegarty and M.J. Topping (1991), Tariq Rahman for discussions on the subject of HANDY 1 | A Low-Cost Robotic Aid to customized design of rehabilitation aids. Eating, Proc. Int'l. Conf. on Rehabilitation Robotics, pp. 17-25. [11] N. Hogan (1982), Control and coordination of voluntary arm movements, Proc. Ameri[1] N.I. Badler, C.B. Phillips and B.L. Webcan Control Conf., pp. 522-528. ber (1993), Simulating Humans: Computer Graphics, Animation, and Control, Oxford [12] J.M. Hollerbach (1989), A survey of kineUniversity Press, New York, NY. matic calibration, Robotics Review I, Eds: O. Khatib, J.J. Craig and T. Lozano-Perez, [2] M.F. Burrow and T.G. Single (1990), The MIT Press, Cambridge, MA, pp. 207-242. Georgia Tech Robotic Manipulator: A Six Degree of Freedom Arm for Rehabilitation [13] I. Kakadiaris, D. Metaxas and R. Bajcsy Applications, Proc. Int'l. Conf. on Rehabili(1994), Active part-decomposition, shape tation Robotics, Wilmington, DE, pp. 41-51. and motion estimation of articulated objects: [3] G. Bush, I. Al-Temen, J. Hancock, J. Bishop, A physics-based approach, Proc. IEEE Conf. E.M. Slack and I. Kurtz (1994), Develon Computer Vision and Pattern Recogniopment of a Myoelectrically Controlled tion, Seattle, WA, June 21-23. pp. 980-984.

References

[14] H. Kazerooni (1990), Human Machine In- [25] R. Pito and R. Bajcsy (1995), A Solution teraction via the Transfer of Power and Into the Next Best View Problem for Autoformation Signals, IEEE Trans. Systems, mated CAD Model Acquisition Using Range Man and Cybernetics, Vol. SMC-20, No. 2, Cameras, Modeling, Simulation, and Control pp. 450-463. Technologies for Manufacturing, SPIE Photonics East, Philadelphia, PA. [15] C.J.J. Paredis and P.K. Khosla (1993), Kinematic Design of Serial Link Manipu- [26] M. Raghavan and B. Roth (1990), Kinematic analysis of the 6-R Manipulator of General lators from Task Speci cations, Int'l. Jrnl. Geometry, Proc. 5th Int'l. Symposium on of Robotics Research, Vol. 12, No. 3, June, Robotics Research, MIT Press, Cambridge, pp. 274-287. MA, pp. 263-269. [16] V. Koivunen (1993), Processing and interpretation of 3-D sensory data with applica- [27] T. Rahman, W. Harwin, S. Chen and R. Mahoney (1994), Rehabilitation robot contion to geometric modeling, Ph.D. dissertatrol with enhanced sensory feedback, Proc. tion, Department of Electrical Engineering, 4th Int'l. Conf. on Rehabilitation Robotics, Universitas Ouluensis, Finland. Wilmington, DE, June 14-16, pp. 43-48. [17] V. Koivunen and R. Bajcsy (1993), Geometric methods for building CAD models from [28] J.B. Reswick (1990), The Moon Over Dubrovnik | A Tale of Worldwide Impact range data, Proc. SPIE Geometric Methods on Persons with Disabilities, Advances in exin Computer Vision II, San Diego, CA. ternal control of human extremities, 4092, [18] V. Kumar, R. Bajcsy, W. Harwin and 1990. P. Harker (1996), Rapid Design and Prototyping of Customized Rehabilitation Aids, [29] J.H. Sheridan (1993), Agile manufacturing: stepping beyond lean production, Industry Communications of the ACM, Special SecWeek, Vol. 242, No. 8, pp. 3-46. tion on Computers in Manufacturing, (to be published). [30] C.A. Stanger, A.C. Phalangas and M.F. Cawley (1994), Range of Head Motion and [19] H.H. Kwee (1990), Rehabilitation Robotics: Force of High Cervical Spinal Cord Injured Softening the Hardware, Proc. Int'l. Conf. Individuals for the Design of a Test-bed on Rehabilitation Robotics, pp. 69-79. Robotic System, Proc. 4th Int'l. Conf. on [20] D. Lees, R. Crigler, H.F.M. Van der Loos and Rehabilitation Robotics, Wilmington, DE, L. Leifer (1988), A Third Generation DeskJune 14-17, pp. 37-42. top Robotic Assistant for the Severely Disabled, Proc. ICAART 88, Montreal, Canada, [31] K.T. Ulrich and S.D. Eppinger (1995), Product Design and Development, McGraw-Hill, June 1988. New York, NY. [21] L. Leifer (1992), RUI: factoring the robot user interface, Proc. RESNA Int'l. '92, [32] G. Verberg, M. Milner, S. Naumann, J. Bishop and O. Sas (1992), An EvaluaRESNA Press. tion of the Manus Wheelchair-mounted Ma[22] \National Spinal Cord Injury Statistical nipulator, Proc. RESNA Int'l. '92, RESNA Center Annual Report," University of AlPress. abama at Birmingham, 1992. [33] J. Vertut and P. Coiet (1986), Teleopera[23] M. Ouerfelli (1994), Identi cation of Open tions and Robotics: Evolution and Developand Closed Kinematic Chains with Apment, Robot Technology, Vol. 3A, pp. 191plications to Robotics and Biomechanics, 194, Prentice Hall. Ph.D. dissertation, Department of Mechanical Engineering, University of Pennsylvania, [34] P. Wellman, V. Krovi, V. Kumar and W. Harwin (1995), A Wheelchair with Legs Philadelphia, PA. for People with Motor Disabilities, IEEE [24] P. Papalambros and D. Wilde (1988), PrinTrans. on Rehabilitation Engineering, (in ciples of Optimal Design, Cambridge Univerpress). sity Press, New York, NY.