Accessible Interfaces for Robot Assistants
Contents
1. computer to them This raises the question if a human assistant has to bring the computer can t they also help with the task rather than having the robot do it Even for persons with normal physical abilities requiring a computer to interact with the robot is a limitation we would like to remove We describe a system that allows a person with limited physical abilities to assign tasks to a complex mobile ma nipulation robot in a home setting For the work reported in this paper we assume that the person has good control of their head position We present a system for use with a variety of household tasks that uses only the motion of a user s head and a single click to interact with task specific interfaces projected into the world or onto relevant objects To do so the system a uses the robot s on board sensors to estimate the head pose of the person b combines this with information from the world models maintained by the robot to generate context sensitive interface elements and c projects these interface elements directly onto the world allowing the person to interact with them using a cursor controlled by the user s head motion Il RELATED WORK Before going on to describe our system in detail we first discuss some of the related work in interfaces and robot control Graphical user interfaces typically rely on some sort of pointing device 7 The mouse or some similar device is by far the most common device althou2. performed by home automation to be used in uninstrumented environments such as public spaces However the question of how best to direct the robot to perform these tasks is still an open one Interfaces to assign tasks to these mobile manipulation robots typically involve either physical gestures interpreted by the robot for example 4 or a custom designed graphical interface displayed on a computer screen for example 5 6 While these approaches have been shown to work well in a number of systems they make some implicit assumptions about the person directing the robot To use physical gestures a person must be able to move their arms To use a graphical interface a person must have a computer in front of them and be able to use it Both of these assumptions limit the usefulness of a robot assistant for persons with severe motor disabilities who otherwise might most benefit from such a system If a person is unable to effectively move their arms they cannot use a gestural interface While many persons with physical disabilities can use a computer through the use of alternative input devices a human assistant often has to bring this Daniel A Lazewatsky and William D Smart are with the School of Mechanical Industrial and Manufacturing Engineering Oregon State University Corvallis OR 97331 USA This work was funded in part by a research gift from Willow Garage Tlazewatd engr oregonstate edu 2bill smart oregonstate edu
3. on Open Source Software 2009 H Kato M Billinghurst and I Poupyrev ARtoolkit user manual version 2 33 2000 Human Interface Technology Lab University of Washington M A Fischler and R C Bolles Random sample consensus A paradigm for model fitting with applications to image analysis and au tomated cartography Communications of the ACM vol 24 pp 381 395 June 1981 G Fanelli T Weise J Gall and L V Gool Real time head pose estimation from consumer depth cameras in 33rd Annual Sympo sium of the German Association for Pattern Recognition DAGM 11 September 2011 D A Lazewatsky and W D Smart Context sensitive in the world interfaces for mobile manipulation robots in Proceedings of the 21st International Symposium on Robot and Human Interactive Communi cation Ro Man pp 989 994 2012 W N Francis and H Kucera The brown corpus A standard corpus of present day edited american english 1979 Brown University Liguistics Department M Ciocarlie K Hsiao E G Jones S Chitta R B Rusu and I A Sucan Towards reliable grasping and manipulation in house hold environments Proceedings of the International Symposium on Experimental Robotics ISER 2010 M Waibel M Beetz J Civera R D Andrea J Elfring D Galvez Lopez K Haussermann R Janssen J M M Montiel A Perzylo B Schiessle M Tenorth O Zweigle and R van de Molengraft Roboear
4. use a more complex non planar world model generated by the robot such as a 3d mesh the only difference would be in the computational cost of the intersection calculations The planar model parameters can be estimated in two ways We can add markers in the form of augmented reality AR tags 22 to the relevant surfaces and use a monocular camera such as a webcam to determine the 3d location and orientation of the tag and hence the surface We can also use the more advanced sensors mounted on the robot which generate clouds of 3d points corresponding to objects in the world A planar model can then be fit to the point cloud using the Random sample consensus RANSAC algorithm 23 This algorithm works by successively selecting a random subset of the data as inliers and testing how well those data fit the given planar model Once a set of inliers has been chosen the algorithm then estimates the model parameters using those inliers C The Pointing Input The system incorporates user input in the form of a 3d vector that points at objects in the world The intersection of this vector and the world model allows us to determine the 3d point in the world that the user is attending to Although this vector can be estimated from a number of input sources for the work reported in this paper we use an estimate of the user s head pose both position and orientation for the pointing input When using a planar world model the p
5. Accessible Interfaces for Robot Assistants Daniel A Lazewatsky and William D Smart Abstract Currently high level task control of robots is generally performed by using a graphical interface on a desktop or laptop computer This type of mediated interaction is not natural and can be problematic and cumbersome for persons with certain types of motor disabilities and for people interacting with the robot when there are no computer displays present In this work we present a framework which enables the removal of such obvious intermediary devices and allows users to assign tasks to robots using interfaces embedded directly in the world by projecting these interfaces directly onto surfaces and objects We describe the implementation of the projected interface framework and give several examples of tasks which can be performed with such an interface I INTRODUCTION The idea of a robot assistant has been around for a long time However recent advances in perception manip ulation and autonomy are bringing this vision closer to reality There are already many tasks that robots can perform autonomously from picking up and delivering household objects 1 to opening doors drawers cabinets and contain ers 2 to cooking complete meals 3 These advances have the potential to make robot assistants truly useful especially for persons who cannot perform some of these actions for themselves They also enable the sorts of tasks often
6. ROS software infrastructure 21 ROS is free and open source and provides a simple and standard way of interacting with sensors and actuators The system comprises three main components a model of the world maintained by the robot a pointing input generated by tracking the user s head pose and a projected interface that allows the user to task the robot We describe each of these four components in turn A The Robot For all interactions requiring a robot we use a Willow Garage PR2 robot PR2 has a quasi holonomic base two 7 degree of freedom arms and a movable head containing a variety of sensors including two pairs of stereo cameras a textured light projector a high resolution camera and a Microsoft Kinect Additionally PR2 has a planar laser rangefinder on the base and another planar laser rangefinder mounted on a tilting platform on the torso which can create 3d world models B The World Model The robot builds and maintains a 3d model of the world with its sensors For the work reported here however we only use part of this model We extract the plane corre sponding to the surface onto which the robot will project the interface This plane is represented by its normal vector and a point on the plane in the robot s coordinate frame Using a simple parameterized model for the plane allows us to perform fast intersection calculations to determine where the user is looking It would be equally easy however to
7. e state machine the state of which depends on whether or not the robot is currently grasping an object When no object is being grasped a click on a valid object directs the robot to pick up that object When an object is being grasped a click anywhere in the workspace directs the robot to put down the object at the indicated location This interface can be augmented with other task dependent elements For example for a sorting task areas can be projected onto the workspace for each category assisting in object placement An example of this is shown in figure 2c with several manipulatable objects circled on a table D Light Switch Users should also be able to control the physical infras tructure in their environments As an example of this we have created an interface that allows the user to turn on and off a light switch The robot first detects and categorizes the light switch and places an interface element that says light switch over it Clicking on this element causes a context sensitive menu to be displayed as shown in figure 2d This menu enumerates all of the physical manipulations that can be performed on the light switch Clicking on turn on for example will cause the robot to move over to the light switch and actuate it with its gripper Once a device is detected and classified it can be stored in the world model maintained by the robot This allows the device to be used in the future without the detectio
8. erfaces can be composed in the real world positioning elements with respect to objects or markers in the world and dimensions can be specified in meaningful units such as meters This makes it easy to design interfaces that fit with the objects they control and ensures that angles and measurements are reproduced accurately for example guaranteeing that ele ments which should be rectilinear are rectilinear regardless of the placement or orientation of the projector with respect to the projection surface B Interaction Styles We are interested in enabling interactions which require a user simply to walk up to the robot to begin interacting with it However these interactions will always be embedded in some context which will allow us to simplify and specialize the interface elements dynamically External context 1s supplied by where and when the interaction takes place The range of robot tasks in the kitchen for example will be different from those in the Fig 3 An example of a letterboard used for alternative communication Users communicate by looking at letter groups sequences of which are interpreted by someone experienced with the system such as a caregiver family or friend dining room and this will let us specialize the interfaces to make the interaction more efficient Similarly the tasks that the user assigns the robot in the morning might be different from those assigned in the evening Since the robot is ca
9. events can trigger a mouse click For example the Kinect sensor we use to estimate head position could detect when the user opens their mouth and use this to initiate a mouse click For the rest of this paper the term mouse click should be taken to mean a discrete signal that the user can send to communicate to the system that the cursor is currently over the object of interest D System Calibration In order to be able to accurately project onto locations in the world and to determine the relationship between head orientation and objects in the world an initial calibration step is required Because we are using a PR2 robot we can assume that all of the sensors and actuators are already calibrated so the only additional calibration step is to find the relationship between 3d world locations and projected pixels This relationship is a homography a linear mapping between pixels in the camera used to model the world and projected pixels denoted by the matrix H H can be found using standard techniques by projecting a known calibration pattern and detecting it with the camera To find H we need at least four points whose locations are known in both the project s pixel coordinates and the camera s pixels coordi nates To project onto any 3d location we can now project the 3d location into pixels in the camera s coordinates and then use H to find the corresponding projector pixels This calibration is very similar to sys
10. gh pen based devices first seen in Sutherland s Sketchpad 8 are relatively widespread However mouse and pen interfaces are often difficult for persons with motor disabilities to use Alternative input devices such as eye tracking 9 or other mouse replacement devices such as TrackerPro R 10 have been developed to emulate mouse input to enable use of standard graphical interfaces There are a few examples of systems that project interface elements into the world SixthSense 11 uses a wearable device incorporating a computer projector and camera to project interactive interfaces onto the world These interfaces can be informational projecting departure gate information onto an airline ticket or functional a working calculator projected onto one s hand While SixthSense is very similar in spirit to the work reported in this paper it lacks the rich sensor information and world models available to our robot and the ability to move around and effect changes in the world independent of its wearer PICOntrol 12 uses a handheld projector and sensor package along with small sensor units on devices to allow users to send simple commands to devices Cao et al use a handheld projector to enable users to explore virtual spaces and interact with virtual objects using a pen and movement of the projector itself 13 Projected interfaces using static projectors have been used with fixed industrial robot arms 14 Again th
11. is is similar in spirit to our system although it is in a fixed location and uses a custom designed handheld interaction device for user input Sasai and colleagues 15 have demon strated a system that projects a control interface for a simple mobile robot onto the floor that allows users to direct the robot using foot taps on the interface This work is similar to ours although it is designed for single type of interaction direction giving and the projection assumes a clear open floor in a known position with respect to the robot Sato and Sakane use a fixed projector and robot arm to project onto a workspace and perform simple pick and place tasks 16 Gesture interfaces tracing their history back to Bolt s Put that There system 17 allow a user to use pointing gestures to interact with an interface Some of these systems such as X Wand 18 use custom interaction devices while others interpret natural human body gestures Gesture based interfaces on robots have enjoyed less suc cess than those aimed at interacting with fixed displays Kemp s Clickable World 19 is a notable exception that uses a standard laser pointer to designate objects for a mobile manipulation robot to fetch Looper and colleagues 20 describe a system that interprets and responds to a limited set of stylized human gestures military hand signals II IMPLEMENTATION Our system is currently implemented on a PR2 mobile ma nipulation robot using the
12. long the way In our reimplementation we use head pose as a proxy for eye gaze Our interface is context free in that it does not rely on any objects or properties of the physical world other than a usable projection surface It can however be made more ef ficient by adding context to interactions by using a language model to perform text prediction We have implemented this by creating a scored set of bigrams from one of the standard linguistics corpora 26 and ordering word suggestions based on their scores from the previous word and current partial word This can be extended even further by learning a language model for each user seeded by for example all of their sent emails and updated as they use the interface The robot detects the wall estimates a parametric planar model of it and projects the interface onto the surface taking this model into account In our interface shown in figure 2a the letters and numbers from the original are presented statically comprising most of the area of the interface A dynamic list of predicted words appears down the right hand side and the current sentence is shown along the bottom In the figure the user is pointing at today with their head pose and this interface element is highlighted in green Clicking the mouse button will select it and add it to the sentence When they are finished clicking on the completed sentence causes the robot to speak it using a standard text to speech
13. ment in Pro ceedings of the Fourteenth International Conference on Technology and Persons with Disabilities CSUN 1999 3 7 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 Madentec Inc Trackerpro http www madentec com products tracker pro php P Mistry and P Maes Sixthsense A wearable gestural interface in ACM SIGGRAPH ASIA 2009 Sketches p 11 ACM 2009 D Schmidt D Molyneaux and X Cao PICOntrol Using a handheld projector for direct control of physical devices through visible light in Proceedings of the 25th Annual ACM Symposium on User Interface Software and Technology UIST pp 379 388 ACM 2012 X Cao and R Balakrishnan Interacting with dynamically defined information spaces using a handheld projector and a pen in Proceed ings of the 19th Annual ACM Symposium on User Interface Software and Technology UIST pp 225 234 ACM 2006 G Reinhart W Vogl and I Kresse A projection based user interface for industrial robots in Proceedings of the IEEE Symposium on Virtual Environments Human Computer Interfaces and Measurement Systems VECIMS pp 67 71 2007 T Sasai Y Takahashi M Kotani and A Nakamura Development of a guide robot interacting with the user using information projection basic system in Proceedings of the Internati
14. n and classification step The locations and types of switches and other infrastructure elements could even be entered into a persistent world model by a human to remove the need for recognition and classification completely VI FUTURE WORK AND DISCUSSION Ultimately we envision these projected interfaces as one piece in a larger system for enabling anyone but especially users with physical disabilities and visual impairments to control mobile robots in a variety of tasks in their homes Central to this will be a rich persistent model of the world where the robot can store information about the environ ment This information can be used to give context to the interactions and will allow us to make interfaces that take advantage of this context Some of the tools necessary for such models already exist such as the ability to build semantic maps which can provide much richer world models than those we have presented Semantic maps store meaningful information about objects and locations which could include data such as locations of light switches or of objects with which the robot knows it can interact or users are interested in interacting with Projects such as RoboEarth 28 could also be leveraged for information about recognizing and interacting with pre viously unknown objects In this paper we presented a framework for embedded interfaces for use with mobile manipulation robots The system is designed to be usable by person
15. off by the user facing the TV and clicking again with the projector on the robot providing a cursor for feedback More complicated functions such as changing channels are possible by creating simple interfaces with buttons for these functions The interfaces Tt ae ss how are you BEF Fig 2 Several example applications a Letterboard interface in the process of saying how are you today b TV interface showing a user selecting channel up c tabletop interface showing several detected objects d an interface for controlling a light switch The Turn On button is highlighted indicating that the switch is currently on can either be projected which will not interfere with normal use of the TV or by using the TV as the interface display device If projected the small controls on the TV remote can be made arbitrarily large within the limits of the projection system and unnecessary controls can be left out affording those with visual impairments improved access to the controls In the interface shown controls for on off and channel up and down are displayed When the TV is off only the on off button is displayed and turning on the TV causes extra controls to be displayed Additional controls can be easily added to the interface and controls can be hidden or displayed based on the state of the TV Accessibility can be further increased by incorporating more task context into the interaction Standard TV
16. oint can be found using the simple plane ray intersection calculation This is only valid if the ray is already known to intersect the plane somewhere and is not contained within parallel to the plane The first condition is satisfied by assuming the projection surface is an infinite plane The second property holds because the camera used to track the user is always pointing approximately away from the projection surface and can only track faces from a frontal view When using a point cloud representation of the world the point is found by intersecting the ray with the point cloud This can be performed efficiently using an octree representation of the point cloud which enables expected O log n ray tracing operations When ray tracing we can return the intersected point closest to the user because any points father away would be occluded from the user s view However this intersection calculation is still slower than the constant time plane ray calculation and scales albeit logarithmically with the size of the world model 1 Head Pose Estimation In the current system head pose estimation is performed in real time using depth data collected from a Microsoft Kinect sensor The estimation is performed using the system described by Fanelli et al 24 This technique takes noisy depth data and produces a 6 degree of freedom pose estimate containing the 3d position of the head as well as the head s orientation an example of which i
17. onal Conference on Mechatronics and Automation PICMA pp 1297 1302 2011 S Sato and S Sakane A human robot interface using an interactive hand pointer that projects a mark in the real work space in Robotics and Automation 2000 Proceedings ICRA 00 IEEE International Conference on vol 1 pp 589 595 vol 1 2000 R A Bolt Put That There Voice and gesture at the graphics interface in Proceedings of the 7th Annual Conference on Computer Graphics and Interactive Techniques SIGGRAPH pp 262 270 ACM 1980 A Wilson and S Shafer Xwand UI for intelligent spaces in Pro ceedings of the SIGCHI Conference on Human factors in Computing Systems pp 545 552 ACM 2003 H Nguyen A Jain C Anderson and C Kemp A clickable world Behavior selection through pointing and context for mobile manipu lation in Proceedings of the IEEE RSJ International Conference on Intelligent Robots and Systems IROS pp 187 793 September 2008 M M Loper N P Koenig S H Chernova C V Jones and O C Jenkins Mobile human robot teaming with environmental tolerance in Proceedings of the 4rd International Conference on Human Robot Interaction HRI pp 157 164 ACM 2009 M Quigley K Conley B Gerkey J Faust T Foote J Leibs R Wheeler and A Y Ng ROS An open source robot operating system in Proceedings of the IEEE International Conference on Robotics and Automation Workshop
18. pable of estimating its position and its physical environment we can use this to infer the appropriate context of the interaction Task context 1s context that can be inferred or learned from the task itself For example if the user always has a particular brand of cereal in the morning the interface can be specialized to place that choice in a prominent location in the interface This preference could either be pre supplied to the robot or potentially learned over time through repeated interactions We can use both the external and the task context to modify the interface presented with the goal of making the interaction as efficient as possible We give some examples of this in the next section V EXAMPLE APPLICATIONS A Letterboard To illustrate a simple interaction with our system we created a projected version of a standard augmentative and alternative communication AAC device a gaze based let terboard The particular letter board shown in figure 3 is one used by a colleague of ours who has quadriplegia and is mute Use of the physical version of this letterboard involves an able bodied listener holding the board between themselves and the AAC user The AAC user spells out words by using eye gaze to indicate letter groups to the listener The listener must correctly identify the letter that the AAC user is looking at and then uses the letter groups to infer what the AAC user is saying asking for confirmation a
19. remotes are dumb devices with a few exceptions They know noth ing about the user or their preferences Since our interface is mediated by a robot that is connected to the Internet we can display interfaces that give program listings show names or other contextually appropriate selection options If we assume that a person will use our system for an extended period of time we can learn or have programmed in their preferences If they always watch the channel 9 news at 9pm we can adjust the interface presentation accordingly since we know the time C Tabletop Manipulation A common task for manipulation robots involves moving objects around on a tabletop This is an important ability for a variety of useful tasks that the robot might perform under the direction of a person Presented in detail in 25 an interface for directing a robot in pick and place tasks can easily be created In this interface all objects with which the robot is able to interact are highlighted by drawing circles around them A cursor representing the point where the ray from the user s head pose intersects the world model is projected onto the work surface which both shows the user where the system believes them to be pointing and also indicates where the robot is able to pick up or place objects Object detection grasp planning and execution 27 are all performed by modules which are core packages within ROS The system runs a simple two state finit
20. rs and M Beetz A generalized framework for opening doors and drawers in kitchen environments in Proceedings of the International Conference on Robotics and Automation ICRA St Paul MN USA May 14 18 2012 M Beetz U Klank I Kresse A Maldonado L Mo6senlechner D Pangercic T R hr and M Tenorth Robotic Roommates Making Pancakes in Proceedings of the llth IEEE RAS International Con ference on Humanoid Robots Bled Slovenia October 26 28 2011 4 S Waldherr S Thrun and R Romero A gesture based interface for human robot interaction Autonomous Robots vol 9 no 2 pp 151 173 2000 5 A Leeper K Hsiao M Ciocarlie L Takayama and D Gossow Strategies for human in the loop robotic grasping in Proceedings of the 3rd International Conference on Human Robot Interaction HRI Boston MA pp 1 8 2012 6 H Nguyen M Ciocarlie K Hsiao and C Kemp Ros commander Flexible behavior creation for home robots in Proceedings of the International Conference on Robotics and Automation ICRA In Press D Engelbart Augmenting human intellect A conceptual framework 2001 8 I E Sutherland Sketchpad A man machine graphical communi cation system in Proceedings of the SHARE Design Automation Workshop pp 6 329 ACM 1964 9 D Rasmusson R Chappell and M Trego Quick glance Eye tracking access to the Windows 95 operating environ
21. s shown in figure 1 Although we use the Kinect sensor for the work reported here any source of 3d point data would work equally well Fig 1 A point cloud view of a user showing the user s head pose estimate as a vector This estimate is quite noisy With the user at approximately lm from the Kinect the standard deviation in the roll pitch and yaw angles was found to be 0 62rad 0 12rad and 1 37rad respectively At 2m from the projection surface this translates to the cursor from a stationary user being within a circle of diameter approximately 9 521cm with 95 confidence This problem only gets worse as the distance to the projection surface increases or the obliqueness of the angle increases We have previously evaluated the Kinect as a pointing device and found that despite the noise novice users are able to effectively use it in object designation tasks 25 2 Mouse Clicks Our system relies on the user being able to perform actions analogous to mouse clicks This can be done with a traditional computer mouse if the user is physically able to operate one well enough to simply click one of the buttons even if they cannot move the mouse on a surface This is sometimes the case even for people with severe motor disabilities It can also be done any one of a variety of augmentative and assistive communication AAC devices such a special purpose switches or sip puff devices If we want to avoid additional hardware other
22. s with severe motor disabilities by using only simple motions as input It additionally removes the need for interactions to be mediated by a traditional personal computer and monitor moving interactions out into the real world We additionally presented several illustrative use cases and discussed how different types of contexts affect the interaction The framework is quite general and will work with any input device that can generate a pointing vector Our implementation uses visual head pose estimation but a laser pointer orientation sensors in Google Glass or some other device could be used with no modifications to the framework As robots become more capable of performing useful work in people s homes the interfaces to support that must become more integrated with the environments in which the tasks take place By moving interfaces from computer screens to the objects to be manipulated themselves we hope we have taken a step in that direction VII ACKNOWLEDGMENTS We would like to thank the members of the Robots for Humanity project 29 both at Willow Garage and at the Healthcare Robotics Lab at Georgia Tech as well as Henry and Jane Evans REFERENCES 1 K Hsiao M Ciocarlie and P Brook Bayesian grasp planning in Proceedings of the International Conference on Robotics and Automation Workshop on Mobile Manipulation Integrating Perception and Manipulation 2011 2 T R hr J Sturm D Pangercic D Creme
23. system We note that this interface is a particularly simplistic virtual version of the physical letterboard Our intention in showing it is only to illustrate a simple use case of our system However even with this simple system the AAC user is able to directly communicate with anyone not just those able to interpret physical letterboard gazes in any location as long as there is a flat surface and they are accompanied by their robot B Television In addition to simple communication interfaces the system can also be used to interact with objects in the real world Many objects already have affordances for changing their State either on the device or on external control devices such as a TV remote control These types of interfaces present two challenges First decoupling the interface from the device requires users to divide their attention 12 Sec ond devices such as remote controls often have an abundance of options which can be difficult even for able bodied users and impossible for persons with motor disabilities or visual impairments Pairing embedded projected interfaces with existing de vice controls has the potential to enable powerful yet simple interactions Using an infrared transceiver module we have built an interface that enables users to control TV functions with simple head movements Since the location of the television is known to the robot in the world model that it maintains the TV can be turned on and
24. tem presented in 16 One advantage that naturally falls out of this type of calibration 1s the elimination of any need for explicit keystone pincushion or any other sort of distortion correction IV INTERACTION METHODS In this section we describe the basic building blocks of our projected interface and how they fit together A Interface Elements All projected interfaces are built from a small set of simple polygonal elements which can be annotated with text see figure 2 for some representative examples A cursor is overlaid on the interface providing the user with feedback on where the system thinks they are pointing If the cursor is within an interface element s selection space the selected element is highlighted to indicate that it is active With an active element a click from the user will change the highlight color to indicate that the click has been received and will dispatch a message to the control software containing the ID of the interface element which the user has selected Additionally if the cursor location is outside of the projectable area a bar is displayed on the edge of the projectable area indicating the direction of the off screen cursor Previous results from 25 indicated that providing feedback in this situation is extremely helpful for users Because the mapping between real world coordinates and projected coordinates is known we have fine control over the geometry of the projected interface Int
25. th Robotics Automation Magazine vol 18 no 2 pp 69 82 2011 T L Chen M Ciocarlie S Cousins P M Grice K Hawkins K Hsiao C C Kemp C H King D A Lazewatsky A E Leeper H Nguyen A Paepcke C Pantofaru W D Smart and L Takayama Robots for humanity Using assistive robotics to empower people with disabilities Robotics and Automation Magazine vol 20 pp 30 39 March 2013
Download Pdf Manuals
Related Search