The RoboCare Assistive Home Robot: Environment, Features and Evaluation

Progetto RoboCare: sistema multi-agenti con componenti fisse e robotiche mobili intelligenti Settore “Piattaforme ITC abilitanti complesse ad oggetti distribuiti” MIUR legge 449/97 per l’anno 2000 The RoboCare Assistive Home Robot: Environment, Features and Evaluation A. Cesta1 , G. Cortellessa1 , M. V. Giuliani1 , L. Iocchi2 , G. R. Leone2 , D. Nardi2 , F. Pecora1 , R. Rasconi1 , M. Scopelliti1 and L. Tiberio1 1 Istituto di Scienze e Tecnologie della Cognizione Consiglio Nazionale delle Ricerche <name>.<surname>@istc.cnr.it 2 Dipartimento di Informatica e Sistemistica Università di Roma “La Sapienza” <surname>@dis.uniroma1.it The RoboCare Technical Reports — RC-TR-0906-6 The RoboCare Assistive Home Robot: Environment, Features and Evaluation Amedeo Cesta1 , Gabriella Cortellessa1 , Maria Vittoria Giuliani1 , Luca Iocchi2 , G. Riccardo Leone2 , Daniele Nardi2 , Federico Pecora1 , Riccardo Rasconi1 , Massimiliano Scopelliti1 and Lorenza Tiberio1 1 Istituto di Scienze e Tecnologie della Cognizione, Consiglio Nazionale delle Ricerche, Via S. Martino della Battaglia 44, I-00185 Rome, Italy — <name>.<surname>@istc.cnr.it 2 Dipartimento di Informatica e Sistemistica, Università di Roma “La Sapienza”, Via Salaria 113, I-00198 Rome, Italy — <surname>@dis.uniroma1.it Abstract Calo project [24, 23] has as its primary goal the development of cognitive systems which are able to reason and learn from experience, respond robustly to contingencies, and which can be told what to do and explain what they are doing. The state-of-the-art in robotics allows now an increasing emphasis on human-robot interaction in general and on social assistive robotics in particular. The emphasis in the latter is to support human users through social rather than physical interaction [12]. RoboCare shares several of the challenges with the above mentioned projects. Indeed RoboCare’s main motivations can be summarized as follows: This paper describes results from the RoboCare project, whose aim is to create assistive intelligent environments for older people. The specific goal of the project has been to synthesize a multi-agent system in which robotic, software and sensory services are integrated to offer cognitive support to the older user at home. The paper describes the technology that has been integrated to create an empowered robot that assists the user in day-to-day activities. After providing some details on the implementation of the integrated system, the paper describes results from a controlled experimentation with human users. The analysis is aimed at understanding the perception of potential users with respect to the services that are currently supported by the assisted environment. 1 “The objective of this project is to build a distributed multi-agent system which provides assistance services for elderly users at home. The agents are a highly heterogeneous collection of fixed and mobile robotic, sensory and problem solving components. The project is centered on obtaining a virtual community of human and artificial agents who cooperate in the continuous management of an enclosed environment.” Toward “Robotically Rich” Assistive Environments The use of intelligent technology for supporting elderly people at home has been addressed in various research projects in the last years, e.g., [29, 28, 17, 30]. Recently, increasing attention is being given The project has involved research groups with differto Cognitive Support Systems. As an example, the ent backgrounds with the goal of investigating how 1 state of the art AI and robotics techniques can be combined to create new domestic services for elderly people [9, 1]. The project has produced a prototype of integrated home environment, called RDE (RoboCare Domestic Environment), composed of a robotic interactive agent, some sensors for continuous monitoring, and additional intelligent systems that store and reason upon knowledge about the assisted elder’s scheduled activities. A multi-agent coordination algorithm guarantees the coherence of the behavior of the whole environment. This provides a functional cohesive which invokes the smart home’s services so as to preserve safeness of the person and provide suggestions. The RDE includes a mobile robotic platform with interaction capabilities. This robot provides an interface between the RDE and the user: indeed, the entire smart home is accessible to the user in the form of an assistive robotic companion. Related Work. Human-robot interaction for socially assistive applications is an emerging research topic that involves several heterogeneous disciplines [12, 35]. The increasing number of specialized interdisciplinary events dedicated to this topic proves the importance of integrating different experiences and competencies to succeed. One of the most important aspects of social assistive robots consists in social interaction between human users and robotic agents. In [31] it is highlighted how observation and behavioral analysis of human-robot social interaction in real environments is necessary in order to take into consideration all the divergent factors pertaining to the design of social robots. In particular, in this work the use of observational studies of human-robot social interaction in open, human-inhabited environments has been proposed as a method to provide useful guidance to designers of social robots as well as to improve the evaluation of their interactive capabilities. Other works have investigated the psychological, social and physiological effects of robots on humans for therapeutic purposes. In [36], authors have applied the mental commit robots to assist activities of elderly people at a day-service center. In [10], authors discuss the area of autism and how mobile robots can play a therapeutic role in the rehabilitation of children with autism. Different work has stressed the distinction between hands-on vs. handsoff assistive tasks. Being interested in the latter, it is worth mentioning the work on assisting older people in a healthcare institution [29, 28], and several works on assisting rehabilitation tasks for cardiac patients, i.e., [11]. Within the RoboCare project previous research [32, 33] showed how people have difficulties in depicting a realistic representation of what an assistive robot can actually do in the domestic environment, showing a strong tendency to overestimate “manipulative” abilities and underestimate robots’ cognitive capabilities. The present work further contributes in this direction. In particular our previous study on human-robot interaction within RoboCare focused on users’ attitudes toward an imaginary robotic agent. The present study, on the other In the spirit described in [12] the RDE is an example of Social Assistive Robot, a concept which can be distinguished from Social Interactive Robot [13] because its main task is to monitor and assist the elder user rather than simply interacting with him/her. Since its beginning, RoboCare has raised numerous challenges. In particular one, also reported in [35], has been paramount in our work: “what are the circumstances in which people accept an assistive robot in their environment?”. Other important questions we have strived to answer (or at least investigated) are “how should an elder user communicate with a robot?”, “should the robot look like a human being?”, and, last but not least, “are robots useful in the domestic environment?”. Our project started with a preliminary study [33, 32, 15] aimed at providing a “best guess” for some of the above questions. This paper comes after three years of development in which we have attempted to realize a prototypical domestic environment equipped with an assistive robot. The aim of the paper is to describe an a-posteriori evaluation of the validity of our choices. In particular, we present experiments aimed at understanding the real perception of older people towards the assistance that this robot is able to offer at the moment. 2 now or wait till after dinner; the request is forwarded to a specialized reasoner which propagates the two scenarios (walk now or walk after dinner) in its temporal representation of the daily schedule, and the result of this deduction is relayed to the assisted in the form of verbal advice (e.g., “if you take a walk now, you will not be able to start dinner before 10:00 pm, and this is in contrast with a medication constraint”). The objective of our prototype is to show how a collection of service-providing and very diverse agents (namely, in our specific case, artificial reasoners, robots and smart sensors) can be integrated into one functionally coherent system which provides more added value than the sum of its parts. The type of elementary services deployed in the RDE mirrors the domotic components that will be available on the market in the near future. In this context, a special focus of RoboCare has been to explore the role of an embodied agent which provides an interface between the assisted person and his or her smart home environment. Our integration effort has yielded an integrated environment which interacts with the assisted person through what we have called a robotic mediator. The base on top of which the robotic mediator is built consists in a Pioneer platform. The mobile platform is equipped with additional sensors, namely a laser range finder, a stereo camera and an omni-directional camera, as well as additional computational resources consisting in two laptops, one for on-board sensor processing and navigation and one for human-robot interaction. The robot is endowed with verbal user interaction skills: speech recognition is achieved with the Sonic speech recognition system (University of Colorado)1 , while speech synthesis is driven by a simple text-to-speech system. We start the description of the robotic mediator from its mobility subsystem, i.e., the specific solutions we have adopted to obtain a robot companion which can robustly and safely navigate in the domestic environment. We then briefly discuss the peripheral components of the RDE, namely the sensory systems which we have deployed (person local- hand, is carried out within an actual RDE, which allowed us to develop more realistic situations where the RoboCare robotic platform acts as a cognitive assistant and helps users in cognitive tasks. Plan of the Paper. The paper starts with a technical description of the RoboCare Domestic Environment which includes the robot, the intelligent analysis performed on the vision sensors, the temporal reasoning service based on scheduling technology and the coordination algorithm which guarantees the functional coherence of the multi-agent environment. Section 3 presents the experimental set up and the result obtained with elder users, while Section 4 analyzes the obtained results. A concluding section ends the paper. 2 The RoboCare Experience: Heterogeneous Ingredients in a Smart Home The RDE is aimed at demonstrating instances in which the coordinated operation of multiple household agents can provide complex support services for the elder user. For instance, suppose the assisted person is in an abnormal posture-location state (e.g., lying down in the kitchen). The intelligent home should recognize this situation and react to the contingency by dispatching the robot to the person’s location. The robot should then ask if all is well, and if necessary sound an alarm. Proactive system intervention may also be triggered by complex symbolic reasoning. A meaningful example: the smart environment detects that the time bounds within which to take a medication are jeopardized by an unusual activity pattern (e.g., the assisted person starts to have lunch very late in the afternoon); as a consequence, the system should verbally alert the assisted person of the possible future inconsistency. An even more advanced form of reasoning-driven interaction could be the following: 1 For details, see cslr.colorado.edu/beginweb/speech_ the assisted person asks the intelligent environment (e.g., verbally) whether he/she should take a walk recognition/sonic.html. 3 which triggers the robot to reach a certain (x, y) position in the environment, and a goto-place(dest) primitive through which the robot can be sent to a particular known destination (such as “the sofa”, or “the lamp”). Clearly, the latter functionality is at a higher level of abstraction than the former, and in our system consists in a naming scheme which associates names to coordinate pairs. Therefore, invoking the goto-place() command will result in a look-up in the location database followed by the appropriate invocation of the goto-XY() functionality. Since the core of the mobility infrastructure comes into play at the goto-XY() invocation level, we here briefly describe the topological path planning algorithm underlying this primitive. Autonomously navigating towards a given coordinate pair in the domestic setting is not a trivial problem. It poses both general problems pertaining autonomous navigation, as well as problems which are unique to the domestic environment. Using complete algorithms to find the topology of the environment (e.g., Voronoi diagrams) is very expensive and, since we have a different map at each cycle, a probabilistic approach is more convenient for the topological path-planner. The most widely used approach that builds a graph representing a roadmap of the environment is the Probabilistic RoadMap (PRM) [19] algorithm. This algorithm works by picking random positions in the configuration space and trying to connect them with a fast local planner. The problem with this algorithm is that it expects as input a map that does not change over time. This requirement cannot be upheld in the domestic environment, where some furniture is frequently moved (e.g., chairs, small tables, etc.) and new objects can clutter the environment semi-permanently. In order to overcome this limitation, we employ an algorithm which combines PRM with Growing Neural Gas (GNG) [14]. GNG is a neural network with unsupervised learning, used to reduce the dimensionality of the input space. In this kind of network, nodes represent symbols and edges represent semantic connections between them; the Hebbian learning rule is used in many approaches to update nodes and create edges between them. Given a system which ization/tracking and posture recognition) as well as a software agent which provides the RDE with temporal reasoning capabilities. As the RDE is achieved by integrating diverse technology, we also briefly describe the overall coordination schema which provides service concertation. 2.1 The Robotic Platform In the following paragraphs we briefly describe the functionalities of navigation, path planning, mapping and localization providing the basis for added-value services which require physical presence. Since the overall aim of this paper is to describe the role of the robotic mediator within the assistive system, we omit some details concerning the aspects related to the mobile platform. These aspects are nonetheless given a high-level description, and pointers to specific technical results in localization and artificial vision are given. A significant part of our research in the early stages of the RoboCare project was dedicated to obtaining a reliable and robust mobility subsystem for the robotic mediator. The results of this research are a set of key mobility services consisting in primitives which can be invoked to make the robot reach any position in the domestic environment. Localization and mapping is the primary requirement for implementing a robust mobile platform in the domestic environment. Underlying the mobility services is a Sampling Importance Resampling (SIR) particle filtering algorithm, which is extensively described in [16]. In particular, SIR is particularly suited for the domestic scenario, in which the map of the environment may change in an unpredictable manner. Indeed, the approach allows to take into account the position of chairs, tables, sofas, or any other object whose position is likely to change over time. Given the capability of localizing itself in the environment, the mobile platform must provide a “gotoplace” service which can be invoked in order to make the robot move robustly from one position in the environment to the other. In particular, the RoboCare robotic platform provides two levels of mobility services: a goto-XY(x, y) function on one hand, 4 there is no fixed role. has a finite set of outputs, applying the Hebbian rule allows for modifying the network in order to strengthen the output in response to the input. Otherwise, given two outputs that are correlated to a given input, it is used to strengthen their correlation. For our concerns, the nodes (symbols) represent locations and the edges the possibility to go from one location to another. In this sense, we can use, together with the Hebbian learning rule, a simple visibility check in order to create a link between two nodes, as PRM does. GNG cannot be straightforwardly used in a robot motion problem, because the topological information is valid only when the graph has reached a state of equilibrium. Adding new edges. Concerning the edges, a node is not connected with all its neighbors, as this would be a redundant representation of the environment. Instead, a Hebbian learning rule is used to connect the two nodes nearest to the current input position. Clearly, also the two edges that connect the new position to two nodes that cannot see each other are added. The low number of nodes in the topological map make it easy to move them to different positions as topology changes. It thus becomes easy to check at each cycle if some edges are no longer valid and can be removed. Overall, DPTM has a number of significant advantages compared with PRM and GNG (see [6]). Specifically, it allows to represent the topology of the environment with merely 1% of the nodes required by PRM. Moreover, the density of nodes is a function of the complexity of each portion of the map, and not uniform (as is the case in GNG) thus providing for a good trade-off between accuracy and relevance of the representation. Moreover, using a DPTM we can extract the topology information of the environment, i.e., each path in the environment can be represented on the DPTM, while the PRM algorithm tries to achieve only the connectivity, eventually losing in the graph some connection existent in the environment. This means that with DPTM we can use some method in order to find the optimal path between two positions, while in general with PRM this cannot be done. Overall, DPTM easily adapts to the topological changes of the environment, making it useful in an environment whose map can change often and without notice (as is the case in the domestic environment, where the user can move objects and clutter the map relatively often). Indeed, DPTM is suited for even more dynamic environments, e.g, where the map is built incrementally during exploration such as the rescue scenario. The Dynamic Probabilistic Topological Map. The algorithm we use for domestic navigation is known as Dynamic Probabilistic Topological Map (DPTM) [6]. It successfully combines PRM and GNG, taking into account the characteristics of the considered environment. There are two main issues in this kind of algorithm: (1) when to add a new node (i.e., a new milestone in the topological representation of the environment), and (2) when to add a new edge between two nodes (i.e., when to consider two nodes as part of a path). Intuitively, the approach implemented in DPTM can be described as follows. Adding new nodes. Since only those nodes that are needed to represent the topology of the environment have to be added, the algorithm does not add a new milestone each time a new position is presented to the network, but only if (1) the position cannot see any other node already in the network, and (2) the introduction of a new node makes it possible to connect two nodes that were already present in the network. If a node does not have to be added, the Hebbian learning rule can be used in order to reduce the error distortion in the set of positions represented by the node. This allows to increment the chance of connecting it with other nodes (as 2.2 Environmental Sensors experimented in [38]). The second criterion is similar to that used in [34], but since in this case A major objective of the RoboCare project was the a node could be a connection between two nodes, integration of different intelligent components that 5 Figure 1: The phases of the PLT service (from left to right, top to bottom): original image, intensity foreground, disparity foreground, plan-view, foreground segmentation, and person segmentation. are deployed not only on board of a mobile robot, but also as “intelligent” sensors in the environment. In particular, we have developed a People Localization and Tracking service2 (PLT) based on a stereo vision sensor, which provides the means to locate the assisted person and other people in the environment. This environmental sensor was deployed at RoboCup@Home 2006 in Bremen in the form of an “intelligent coat-hanger”, demonstrating easy setup and general applicability of vision-based systems for in-door applications. The system is scalable as multiple cameras can be used to improve area coverage and precision. In addition, vision-based posture recognition can be cascaded to the PLT computation in order to provide further information on what the assisted person is doing. Our stereo-vision based tracking system is composed of three fundamental modules: (1) background modelling, background subtraction and foreground segmentation, that are used to detect foreground people and objects to be tracked; (2) plan-view analysis, 2 See that is used to refine foreground segmentation and to compute observations for tracking; (3) tracking, that tracks observations over time maintaining association between tracks and tracked people (or objects). The PLT service is effectively capable of tracking the position of a human being within a domestic environment. In addition, the system is resilient to changes in the lighting conditions of the environment, thus enabling portability and easy setup (as demonstrated at the RoboCup@Home competition). This characteristic is particularly useful in domestic environments, where strong differences may occur due to artificial and natural lighting conditions. The key solutions which have made these features possible are: 1. The background model, which is a composition of intensity, disparity and edge information; it uses a learning factor that varies over time and is different for each pixel in order to adaptively and selectively update the model; moreover, it uses a new notion of activity based on edge variations. 2. Plan-view projection computes height maps, which are used to detect people in the environment and refine foreground segmentation in case http://www.dis.uniroma1.it/~iocchi/PLT for an overview. 6 2.3 of partial occlusions. Intelligent Software Services The sensory capabilities provided by the PLT and PPR services are employed in RoboCare to monitor the user’s daily activities. While the above services cannot provide an “all-knowing” smart home, they are sufficient to recognize key activities that are carried out by the assisted person during the day. Activity recognition thus provides a means to assess whether the assisted person is following given behavioral constraints (such as taking the right pills at the right time, eating regularly, and so on). In order to provide the system with the capability to perform these deductions, we have developed a schedule management environment called T-Rex (Tool for schedule Representation and execution [26]), through which it is possible to represent of a set of activities and their quantitative temporal connections (i.e., a schedule of activities that the user is expected to carry out). The broad idea is to allow the specification and then the execution monitoring of a set of activities that the person usually performs or needs to perform due to prescription (on suggestion from his personal doctor for example). T-Rex integrates a constraint-based scheduler [7] with additional features for knowledge engineering, in particular for problem modeling. Particular attention has been given to the definition of “useroriented terminologies” in order to easily synthesize both the basic elements of a target domain as well as different problem instances in the particular domain (i.e., in the form of activities and constraints among activities). For example, in the RoboCare context, T-Rex allows to define “home domain activities” like breakfast, lunch, go-for-walk, and also temporal/causal links among activities like meal-boundmedication to express rules like “aspirin cannot be taken before eating”. Through this high level language an external user (a doctor, a relative of the assisted person, etc.) may define a network of activities, a schedule, that the observed person is supposed to carry out in the home environment during the day. This schedule is dispatched for execution and monitored using the underlying schedule execution technology. Information coming from the sensors is used for maintaining an updated representation of 3. Plan-view positions and appearance models are integrated in the tracker and an optimization problem is solved in order to determine the best matching between the observations and the current status of the tracker. The output of these three phases of the computation is depicted in figure 2.2. In addition to the PLT service, the system also provides a People Posture Recognition (PPR) service. Specifically, this module is cascaded to the PLT module, as its input is the person-blob obtained by the PLT algorithm. In addition, the service relies on a 3D human body model which has been carefully chosen by considering the quality of data available from the segmentation steps. In our application the input data are not sufficient to cope with hands and arm movement. This is because arms are often missed by the segmentation process, and noises may appear as arms. Without taking into account arms and hands in the model, it is not possible to retrieve information about hand gestures, but is still possible to detect most of the information that allows to distinguish among the principal postures, such as STANDING, SITTING, BENT, KNEELING, and LAYING. Our application is mainly interested in classifying these main postures and thus we adopted a model that does not contain explicitly arms and hands. A detailed description of the PLT and PPR services is outside the scope of this paper, and the interested reader is referred to [2, 8] for further descriptions of the technology underlying the PLT and PPR services. Nevertheless, we should underscore that these services are key enabling factors for the sophisticated cognitive support services provided by the smart home. Constant tracking and posture recognition allows to deduce the state of the assisted person, and is therefore responsible for activity recognition. As we briefly explain in the next sections, recognized activities are propagated within a temporal representation of the assisted person’s daily schedule, which in turn triggers the proactive behavior of the robotic mediator (in the form of suggestions, warnings, and so on). 7 what is really happening in the environment. Even if human activity recognition3 is outside the scope of the project, it is worth highlighting how the sequence of observations from the artificial vision sensors allows to follow the evolution of the activities of the observed person (e.g., if and when she took a pill, when she had lunch, etc.). Based on the synthesis of these observations, the system is able to generate a report for the external users that underscores when the person’s activities have been performed within “reasonable” temporal boundaries or when important anomalies or even violations on their execution have been detected [1]. In this light, the RDE constitutes a basic example of home activity monitor grounded on scheduling technology. Notice that, on its own, the domestic activity monitor acts as a “silent observer” and does not take initiative with respect to the elder person in any way. In order to close the loop, we need to show how its indications are employed to trigger system initiatives through the robotic mediator. This is achieved through a distributed coordination infrastructure. A brief description of how this works is given in the following section. Overall, the activity monitor is tightly connected to the interaction between the robotic mediator and the assisted person. Its temporal deductions can give rise to instances of communication with the elder user: if the system recognizes that some temporal constraint is violated, such as taking medication on an empty stomach, the robot will pro-actively intervene by navigating towards the user and communicating a warning message. Also, the temporal reasoning services are exploited when the user spontaneously asks the robot about activity-related concepts, such as “have I taken my pills?” or “when should I start cooking?”. Finally, we should mention that significant efforts have gone into establishing a scheme for deducing verbal messages that are as convincing as possible. In practise, when a temporal constraint is violated, the activity monitor signals the nature of the violation in terms of the two activities involved in the violated constraint. This information is sufficient to distinguish classes of warnings (e.g., something has occurred to early/late, the user is trying to perform activity A before activity B, etc.). These classes correspond to general purpose phrases, and the occurrence of a specific constraint violation thus results in a human intelligible phrase such as “Jane, you should not take your pills on an empty stomach – how about some breakfast?”. 2.4 Multi Agent Coordination Infrastructure RoboCare requires the combination of various intelligent tools to ensure a comprehensive behavior of the enhanced physical environment. Our goal is to achieve an environment which acts as a proactive assistant for daily activity monitoring. This section explains how heterogeneity is kept under control by synthesizing a coordinated behavior. Coordination of multiple services is achieved by solving a Multi-Agent Coordination problem. This problem is cast as a Distributed Constraint Optimization Problem (DCOP), and solved by AdoptN [25], an extension of the Adopt (Asynchronous Distributed Optimization) algorithm [22] for dealing with n-ary constraints. Figure 2 gives an intuition of the approach4 . Let Applicationi be the generic intelligent subsystem that is to be integrated in the overall multi-agent system, and V arj one out of a set V of variables in terms of which the coordination problem is defined. Each variable has an associated domain of Values Dj . Variables are bound by constraints like in regular Constraint Satisfaction Problems (CSP). Conversely, while constraints in CSP evaluate to satisfied or unsatisfied, in the optimization case constraints evaluate to costs. Constraints may involve an arbitrary subset of the variables (n-ary constraints): a constraint among the set C ⊂ V of k variables is expressed as a function in the form fC : D1 × . . . × Dk → N. For instance, a constraint involving the three variables {V ar1 , V ar3 , V ar7 } may prescribe that the cost of a particular assignment of values to these variables 3 There is plenty of recent research on activity recognition with sensors, e.g., [27], that could be potentially impact on this class of applications. 4 8 Further details are given in [8]. wards aggregate behavior which is helpful for the assisted person. amounts to c, e.g., fV ar1 ,V ar3 ,V ar7 (0, 3, 1) = c. The objective of a constraint optimization algorithm is to calculate an assignment A of values to variables while P minimizing the cost of the assignment C∈C fC (A), where each fC is of arity |C|. In RoboCare, the valued constraints are decided in order to orient the solution of the DCOP toward preferred world situations (broadly speaking those situations in which the person is maximally helped by the intelligent system). The system is composed of a number of heterogeneous applications: (a) the TRex activity monitor (described in the previous section), (b) the dialogue manager plus the speech I/O modules, (c) the mobile robotic platform, (d) one application for each of the cameras, each of them with the appropriate software for PLT and PPR. 3 Experiments Users with Elder The RDE’s fundamental building blocks described in the previous sections are the result of a multidisciplinary research and development effort, combining robotics, artificial vision, automated scheduling and distributed constraint reasoning. In all these fields, research has been driven by the specific requirements of the assistive environment scenario. Our aim in the remainder of this article is to provide an evaluation of the validity of our choices. In particular, we present experiments aimed at understanding the real perception of older people towards the assistance that the robot (and thus the assistive environment as a whole) is able to offer at the moment. Another result which has driven our development effort is an a-priori evaluation of laypeople’s perception of assistive robots. Specifically, the study was Figure 2: DCOP to maintain distributed coherence. based on an imaginary assistive robot, and was perEach application manages one or more of the vari- formed before the development of the RDE. This piables which are part of the DCOP. A variable may lot study was aimed at drawing some preliminary represent (a part of) the input to an application, its desiderata and requirements for the RDE. output, or both (see the dashed lines in Figure 2 as an example). When the DCOP resolution algo- 3.1 Preliminary Evaluation of Assisrithm (Adopt-N) is called into play, the values of the tive Robots application-output variables are taken into account in the distributed resolution. When resolution reaches The pilot study, reported in [33], was aimed at anaa fixed-point, the input variables will reflect an up- lyzing laypeople’s representations of domestic robots dated input for the applications. The correspondence with respect to a variety of topics: the respondants’ between the applications’ output at the i-th iteration expectations with respect to the robot’s capabilities and their input at iteration i+1 is a result of the prop- to perform different everyday activities at home; their agation rules specified in the DCOP. Overall the de- emotional response to a domestic robot; the image of cisions of the applications constitute the input for the the robot, referring to shape, size, color, cover masubsequent iteration of the cycle hDCOP-resolution; terial, speed; preferences and expectancies about the read-variable; application-decision; write-variablei. robot’s personification (given name, etc.) and the It is worth underscoring that the multi-agent solu- modalities of human-robot communication and intertion based on DCOP guarantees continuous control action. Results showed that people have difficulties in deover the whole environment. Additionally, the value functions fC allow to bias the produced solution to- picting a realistic representation of what a domestic 9 robot can actually do in the domestic environment, showing a strong tendency to overestimate “manipulative” abilities and underestimate robots’ cognitive capabilities. This is presumably the consequence of the most widespread source of information about robots, namely science fiction: a domestic robot is still too far away from the everyday life experience of laypeople. In addition, people at different stages of their lifespan showed very divergent opinions and preferences about the robot’s appearance and interaction modalities. In particular, older people emerged as a quite homogeneous group in indicating a preference for a small device, with a standard aspect and no sign of personalization, not autonomously free to move in the domestic environment and simply responding to tasks to be performed — a mere task executor, hardly resembling a human being, which has to intrude as less as possible in personal and domestic life. In fact, while its practical utility was recognized, the robot emerged as a potential source of danger and discomfort in private life, and the idea of a non-autonomous device which does not show human features seemed to be a way to ward off their anxiety. Another issue to be addressed has to do with the context in which the robot is expected to operate. The use of new technologies and domestic robots in the home environment is not only a matter of general human-technology interaction, but is also associated with the specific sphere of human life in which assistance is needed [15]. Elderly people showed a rather positive attitude towards a technological modification in the domestic environment, yet the inclination to use technological devices is strongly associated to the problem they have to cope with. In some situations, a technological aid seemed to be unrealistic, or unpractical, or it would have better been replaced by a more common alternative. In other ones, concerning health and personal/environmental safeness above all, it emerged as a suitable solution to cope with losses imposed by ageing. As a consequence, the possibility of identifying everyday activities for which the acceptability of a technological help is likely to be pervasive is undoubtedly an interesting research issue. 3.2 The Present Study The pilot study mentioned above focused on the study of users’ attitudes toward a purely imaginary robotic agent, with no specific abilities and not operating in a real domestic environment. For this reason, differences in users’ reactions could have been related to both diverse knowledge and bias toward technologies. Nevertheless, these preliminary results were important for driving development in RoboCare. The final prototype we have described in this paper allows us overcome the previous limitation. The evaluation of a tangible robot – which is the result of three years of development – allows us to eliminate the pre-concepts and other biases. Performing the evaluation on the RDE prototype allows us to draw specific conclusions on the RDE, but also to concretely answer some general questions relative to the challenges of assistive technology for elderly people. This analysis is in line with current recommendations for the evaluation of complex assistive technology. For instance, it is recognized in [18] that human-robot interaction is to be evaluated on socio-culturally constituted activities outside the design laboratory. In this light, the aim of our research is to analyze the potential reactions of final users to real life interactions between elderly people and an assistive robot. The present analysis considered eight different scenarios, which were meant to be representative of daily situations in which elderly people may be involved. The situations were selected with reference to previous research on this topic [15], ranging from the most emotionally involving to less critical and emotionally neutral, with the aim of exploring elderly people’s evaluations of the potential role of a domestic robot as a useful support to ageing people. Specifically, the study focuses on three aspects. First, we perform an evaluation of how meaningful each scenario is with respect to the respondents’ every day life. This allows us to understand how useful state-of-the-art assistive technology can be in real situations. Moreover, it provides a precious indication as to whether we are employing this technology to solve real needs. Finally, assessing scenario meaningfulness is aimed at understanding the weight we should give to the user’s 10 evaluations in each scenario. Second, we focus on the respondents assessment of our robotic mediator. The RoboCare project has allowed us to perform an evaluation on a real platform within a fully implemented domestic assistive environment. As a consequence, the evaluation is presumably not affected by prejudice and/or knowledge in the area of robotic systems. The analysis focuses on aspects related to the physical aspect of the robot, its interaction capabilities, and in general its suitability in the domestic context (e.g., size, mobility, integration with the environment). In addition, the usefulness of the system is evaluated in the different scenarios. Finally, we observe user preferences with respect to the robot’s resemblance to a human being. Although our robot is not anthropomorphic, it is possible to deploy it in two slightly different versions: one in which the robot has a 3D facial representation (whose lip movement is synchronized with the speech synthesizer), and one without a facial representation. These variants were used to toggle the variable “Resemblance to human beings”, which is emerging as a key component in elderly people’s representation of domestic robots [33]. A final aim was to analyze the influence of age, familiarity with technologies and perceived health conditions on the evaluation of the robotic agent, being all of these variables strongly related to the possibility of elderly people to adopt a technological solution in the management of everyday difficulties [15]. ations in which the robot provides cognitive support to the elderly person. The following topics, referring to critical areas for the elderly, as highlighted by previous research, were considered: (a) management of personal/environmental safety, (b) healthcare, (c) reminding events/deadlines, (d) support to activity planning, (e) suggestions. In the following, the eight scenarios are shortly described. (a) (b) Figure 3: The two versions of the robot corresponding to the different experimental conditions. Scenario 1 [Environmental safety] The actor/actress is sitting on the sofa, watching TV. In the meantime, in the kitchen the sauce on the stove is overcooking. The sensors communicate this information to the robot. As a consequence, the robot moves Materials. toward the actor/actress and says: “The pot is burnEight short movies (ranging from about 30 seconds ing. You should turn it off”. The actor/actress imto little more than one minute) were developed show- mediately goes to the kitchen and turns the stove off. ing potential interaction scenarios between an elderly person and the RDE’s robotic agent in a real domes- Scenario 2 [Personal safety] The actor/actress tic environment. The same scenarios were shot with is sitting on the sofa, reading a magazine. Suddenly, an actor and an actress. The features of the robotic he/she feels ill, and loses consciousness (or faints?). agent were manipulated according to two different ex- The camera recognizes the situation and communiperimental conditions: in the first condition (“Face”) cates this information to the robot. The robot apa robot showing a human speaking face on a note- proaches the actor/actress and says: “Are you all book monitor; in the second (“No-face”), a robot right?”. As it gets no answer, the robot calls the with no anthropomorphic characteristic (see Figure actor’s/actress’ son at work, who calls the medical 3). The eight scenarios presented everyday life situ- emergency. The final scene shows the son and the 11 doctor in the living room with the actor/actress, who feels fine. robot enters the living room and says: “You have been spending all the day at home. Why don’t you go out and have a walk?”. The actor/actress answers: “I reScenario 3 [Finding objects] The actor/actress is ally don’t feel like it... I think I’ll go water the plants sitting on the sofa, and takes a magazine to read. in the garden”. Suddenly, he/she realizes that the glasses are not on events] The acthe table in front of him/her. The actor/actress calls Scenario 8 [Reminding the robot and asks: “Where are my glasses?”. The tor/actress is having breakfast in the kitchen. sensors in the rooms search for the glasses, and fi- The robot reminds him/her: “Today it’s your friend nally find them in the kitchen. The robot answers: Giovanni’s birthday. Remember to call him”. The “The glasses are on the table in the kitchen”. The ac- actor/actress answers: “You are right. I will do it tor/actress goes to the kitchen and takes the glasses, in a while”. Then he/she goes to the living room and then goes back to the sofa and starts reading the mag- calls Giovanni. azine. Tools. Scenario 4 [Reminding analyses] The actor/actress is in the kitchen. He/she is about to have A questionnaire was developed for data collection. It breakfast. When he/she puts the pot on the stove consisted of four sections, plus a final part for socioto warm up the milk, the robot says: “You cannot demographics. The four sections were arranged as have breakfast now. You have an appointment for follows: a medical analysis”. The actor/actress answers: Section 1. Eight fill-in papers, each of them refer“You’re right. I had forgotten all about it!”. ring to one of the eight scenarios, were presented. For each scenario, questions about the likelihood Scenario 5 [Activity planning] The actor/actress of the situation for the elderly person, the utility is having a call in the living room. He/she is speakand acceptability of the robot were asked. ing to the secretary of a clinical center to have an appointment for a medical examination. The secreSection 2. An attitude scale, consisting of 45 tary proposes an appointment for the next day, with Likert-type items, referring to the physical astwo alternatives: one in the morning, the other in the pect of the robot, its behavior and communicaafternoon. The actor/actress asks the robot for evention modalities; the level of integration with the tual engagements in the following day. The robot andomestic environment; the degree of perceived swers: “You have another engagement in the mornintrusion/disturbance of the robot in everyday ing. In the afternoon, you do not have any appointlife and routines; the personal advantages and ment”. The actor/actress accepts the appointment in disadvantages of having such a device at home. the afternoon. Section 3. An emotional scale, consisting of sixteen Scenario 6 [Reminding medication] The actor/adjectives through which respondents have to actress is sleeping on the sofa, and suddenly wakes evaluate the possible presence of the robot in up. He/she does not realize what time is it, and their home. thus he/she asks the robot. The robot answers: “It is four o’clock”. The actor/actress does not remem- Section 4. Questions about familiarity with new ber whether or not he/she took his/her medicine after technologies, worries with cognitive impairments lunch, and asks the robot. The robot answers: “Yes, related to ageing, and perceived health condiyou took it.” tions were asked. Scenario 7 [Suggestions] The actor/actress is The questionnaire consisted of both multiple-choice watching TV on the sofa. It is five o’clock. The and 5-step Likert-type items, to which respondents 12 • understanding the preferences of elder users with respect to the robot’s resemblance to a human being. had to express their level of agreement/disagreement on a scale ranging from 0 (“I totally disagree”) to 4 (“I completely agree”). Participants. Preliminary Analysis. Subjects recruited for this exploratory study were forty elderly people (aged 56-88 years; mean age = 70.3 years). Participants were 13 males and 27 females; as for their educational level, 17.9% attended primary school, 43.6% attended middle school, 25.6% attended high school, 12.9% have a degree. Most of them (82.5%) are retired. Before retirement, 22.5% were teachers, 15% were office workers. A preliminary analysis shows that the selection of scenarios was effective in identifying typical everyday situations. Specifically, the analysis reveals that Personal safety (M = 2.80, sd = .72), Finding objects (M = 2.80, sd = .94), Reminding medication (M = 2.78, sd = .92), and Environmental safety (M = 2.68, sd = .83) are the most likely situations. The Suggestion scenario (M = 1.78, sd = 1.19) was evaluated as somewhat unlikely (see Figure 4). Procedure. Subjects were randomly assigned to one of the two experimental conditions (Face/No-face). The movies were either projected on a notebook monitor, in a face-to-face administration, or on a larger screen, in a small-group administration. All administrations were performed in quiet settings. Two different sequences of presentation of scenarios were used, in order to avoid the potential influence of an order effect of episodes on results. After the vision of each scenario, participants were asked to fill the paper specifically Figure 4: Likelihood of situations. referring to it (Section 1 of the questionnaire). At the end of the whole presentation, subjects were asked to give general evaluations of the robot (Sections 2-4 of the questionnaire), and to fill the final part of the General evaluation of the robot. questionnaire, referring to socio-demographics. As to the different characteristics of the robot (see Section 2 of the questionnaire), its interaction behavior and communication modalities were on average 3.3 Results positively assessed: elderly people like a face-to-face As mentioned, the analysis of the results is aimed interaction with the robot (M = 2.60, sd = 1.23) at investigating four specific aspects related to the and its pace in the domestic environment (M = 2.52, acceptability of a domestic robotic assistant. Specifsd = 1.20); the speech is appreciated as a way to ically, our results are aimed at foster interaction (M = 3.20, sd = .99), the speed • measuring how meaningful each of the eight sce- of voice is adequate (M = 2.67, sd = 1.14) and the narios is with respect to the respondents’ every robot is easy to understand on the whole (M = 3.08, sd = 1.10). The robot’s integration with the home day life; environment is good: elderly people are positively • providing a general evaluation of the robotic me- impressed by its ability to move in the domestic endiator, as well as a specific evaluation of user vironment (M = 3.48, sd = .75), and are not afraid of preferences in the various scenarios; possible damages (M = 1.60, sd = 1.35), even though 13 a total freedom of movement at home is not completely appreciated (M = 1.52, sd = 1.38). Elderly people realize a variety of advantages associated with the presence of the robot in the domestic environment: feeling safer for people living alone (M = 3.23, sd = 1.14), a valid support for cognitive functioning (M = 3.23, sd = .92) and, in general, in the organization of everyday activities (M = 2.98, sd = 1.03); on the other hand, some troubles with the management of the device (repairs, etc.) (M = 2.95, sd = 1.11) and the possible economic costs (M = 3.25, sd = .84) are expected. Finally, the robot is hardly perceived as a potential source of intrusion/disturbance in their personal life (M = 1.43, sd = 1.39) and the possibility for it to take decisions autonomously is highly valued (M = 2.88, sd = 1.30). Scenario-specific evaluation of the robot. A global picture of the robotic mediator reveals a rather positive perception, especially when considering the potential advantages in the management of everyday situations. The robot emerged as a very useful device for Personal (M = 3.10, sd = 1.01) and Environmental safety (M = 2.83, sd = .90), Reminding medications (M = 2.68, sd = .97), and Finding objects (M = 2.63, sd = .98); conversely, not particularly useful in case of Suggestions (M = 1.85, sd = 1.14) (see Figure 5). In addition to utility, the robot was also indicated as a solution users would accept when difficulties arise, again with specific reference to Personal (M = 2.95, sd = 1.06) and Environmental safety (M = 2.55, sd = 1.01). Resemblance to human beings. With respect to the physical characteristics, the robot appears to be slightly pleasant (M = 2.18, sd = 1.38); as to this issue, however, our manipulation emerged to be effective, being the No-face version significantly preferred on the whole (F(1,38) = 6.34, p < .05), specifically appearing both less mechanical (F(1,38) = 5.11, p < .05) and less cold (F(1,38) = 7.25, p < .05). The No-face version was also evaluated as having a significantly higher level of integration with the domestic environment (F(1,38) = 5.65, p < .05) and a larger variety of advantages than the Face version, referring to ease of use (F(1,38) = 9.36, p < .01) and a low need for repair (F (1, 38) = 4.33, p < .05) above all. No significant differences between the two versions, with reference to communication modalities (F (1, 38) = 1.65, n.s.) and perceived intrusion/ disturbance (F(1,38) = 1.55, n.s.) emerged. Finally, the emotional reaction (see Section 3 of the questionnaire) of elderly people to the robot was very good, scoring high the positive adjectives useful (M = 2.90, sd = 1.10), interesting (M = 2.51, sd = 1.30), and relaxing (M = 2.38, sd = 1.14), and being very low the negative adjectives scary (M = .77, sd = 1.01), overwhelming (M = .97, sd = 1.40), gloomy (M = 1.00, sd = 1.36), dangerous (M = 1.05, sd = 1.23), out of control (M = 1.10, sd = 1.14). The only significant difference between the two experimental conditions was referring to the adjective amusing (F(1,37) = 5.93, p < .05), whose score was higher in the No-face condition. In addition, elderly people seemed to be more likely to develop a psychological attachment towards the No-face version than towards the Face version (χ2 = 6.11, df = 2, p < .05). Additional evaluation. Figure 5: We have also analyzed the influence of other variables with respect to user preferences. In order to measure the influence of age, familiarity with technologies and perceived health conditions on the evaluation of the robot, the above variables were analyzed by splitting respondents in two subgroups. A better evaluation of the robot’s integration Utility of the domestic robot for everyday situations. 14 in the domestic environment by the younger elderly (up to 69 years old) than the older elderly (70 years old and more) emerged. This difference shows a tendency to significance (F(1,38) = 3.66, p < .08). As to interaction modalities, the older elderly feel significantly more uncomfortable when interacting with a non-human agent (F(1,38) = 7.88, p < .01). With respect to disadvantages associated to the presence of the robot, they were more afraid to have difficulties in using the robot (F(1,38) = 4.26, p < .05) and in managing robot maintenance (F(1,38) = 4.33, p < .05). The familiarity with technology did not show any significant influence on the robot’s evaluation. Finally, elderly people with better perceived health conditions showed a fondness for teaching the robot how to perform tasks (F(1,38) = 7.89, p < .01), a higher perception of integration with the home environment (F(1,38) = 6.07, p < .05), and a greater ease of use (F(1,38) = 5.66, p < .05) of the robotic agent. Elderly people who evaluated their health conditions less well, expressed a stronger preference for a robot being inert when not engaged in a domestic task (F(1,38) = 4.12, p < .05), not autonomous in both taking decisions (F(1,38) = 4.64, p < .05) and in giving suggestions (F(1,38) = 4.68, p < .05). 4 Discussion The study yielded a variety of interesting results, which can help shed light on possible future developments in research on the interaction between elderly people and domestic robots. Also, this study addresses some general acceptability requirements for robotic agents. The evaluation of eight specific scenarios helps to single out the main concerns in the domestic experience of elderly people. Everyday life is scheduled across a variety of activities, but only some of them are considered of great importance. The management of personal and environmental safety was perceived as a very likely situation at home, which can become a problem when age increases. At the same time, the cognitive impairment associated with ageing can frequently entail difficulties in remembering to do things as well as what has been already done. These cog- nitive weaknesses are crucial especially when related to healthcare. For these activities, elderly people express a positive attitude towards the support of a robotic agent, which is perceived as a useful and appropriate device to overcome difficulties. With respect to other activities, which are not considered to be essential in everyday life, elderly people show a tendency to underestimate the likelihood of occurrence. The results in this case also show a diminished perception of the robot’s potential utility. The general framework outlined in this study is in line with the model of successful aging put forward by [3], which stresses the role of selection and optimization of activities with increasing age, and the importance of compensation strategies to manage the loss of personal resources. The overall evaluation of the robot highlighted a positive reaction towards a variety of specific characteristics. First of all, the robot is appreciated for its ability to communicate. Verbal interaction is highly valued and the interaction modalities involving a face-to-face relationship seemed to reduce a feeling of emotional distance from this device. Second, elderly people showed no manifest apprehension with respect to the integration of the robot in their home, appreciating the safeness of its ability to move and to avoid objects and obstacles. Nonetheless, the issue of safety emerged to be, again, a central concern in their experience, and they would like the robot to move in the domestic environment only when a specific task has to be performed. Third, the idea of the robot as a possible source of intrusion/disturbance in personal life, as depicted in previous research (see [32]) was not outlined: this underlines the difference between studies on mere representations, which may be biased in some way by the availability of examples of unrealistic robots, and research focusing on actual interactions, thus confirming the validity of our approach. The practical benefits associated with with the assistive robot were clearly recognized. The robot can help people in the management of everyday activities requiring an efficient cognitive functioning, which is likely to be defective with increasing age. Above all, our results show a general enhancement in the psychological tranquillity in facing everyday situations, 15 in that the robot can make elderly people feel safer, especially when they live alone. However, elderly people also showed to be aware of possible difficulties with the robot, which are mainly associated with its cost: a key concern for the price they have to pay, both to acquire the assistive robot and to maintain it clearly emerged. The global picture outlined by these results is undoubtedly positive, and it is further corroborated by the analysis of the emotional characterization of the robot. The robotic agent was positively depicted in terms of utility, relaxation and interest, and hardly recognized as a source of danger, fear and other negative affects. A key role in promoting the acceptability of robotic agents is played by their impact on the behavior of the final user, his/her habits and routines, and on the dynamics of interaction with the home environment. This is because continuity in place experience is an essential factor in preserving personal identity [5]. With respect to this issue, the physical aspect of the robot emerged to be an important feature which can help support acceptability. Any allusion to human beings seemed to have an impact on the relationships between elderly people and their domestic environment. In particular, the No-face version of the robot was definitely preferred, and interestingly, the physical aspect proved to affect also the evaluation of other features which are apparently unrelated. In fact, the No-face version was not only perceived as less artificial and psychologically distant from the user, but also better integrated in the home setting and easier to manage. In other words, the better the aspect, the stronger the perception of positive qualities attributed to the robot. Differential judgements were outlined with respect to socio-demographic and psychological characteristics of users, thus confirming previous studies [15]. Age emerged as an important factor in shaping the tendency for elderly people to feel comfortable with a domestic robot. The older elderly perceive greater difficulties in the interaction with and management of the robot. In addition, they show a higher concern towards possible problems in the integration of this device in the home environment. On the whole, they seem to be afraid, at least to some extent, of extreme modifications in their everyday environ- ment [15], which in turn may lead to difficulties of adaptation [39]. Interestingly, familiarity with technologies did not show to affect the robot’s evaluation, suggesting that acceptability for the elderly is not just a matter of frequency of use, but is presumably related to the perception of control on the environment when resources tend to decline. An interesting picture in this light was given by the analysis of the influence of perceived health conditions on the robot’s evaluation. Elderly people in bad health conditions seemed to be mainly concerned with the potential difficulties and risks associated with the presence of the robot in the home, instead of considering the possible support in performing a variety of activities. In other words, the robot might be a further worry adding to their personal health concerns. Conversely, elderly people in good health conditions were more confident in the possibility of controlling negative aspects of this device, and were more aware of the benefits it can provide. For those people, the perception of good health turned into a stronger perception of everyday competence [37, 20] and self-efficacy [21, 4]. Overall, our study has shown how the availability of a situated prototype can greatly enhance the resolution of psychological evaluation. Our findings can be considered an intriguing starting point to address the issue of acceptability of robotic agents in the everyday life of elderly people. 5 Conclusions In this paper we have presented the RoboCare Domestic Environment, an intelligent domotic environment equipped with an assistive robotic companion aimed at providing cognitive support for elder users who wish to maintain their independence. The smart home is the result of the integration of state-of-theart robotic and software agent technology. After presenting the main components of the system, the article focuses on an experimentation aimed at validating the current capabilities of the environment with respect to the expectations of elder users. A key feature of the assistive environment is an autonomous robot which acts as the primary interface between the cognitive support services provided 16 by the multi-agent system and the assisted elder. Through the robot, the domotic services provide active surveillance of the elderly user at home. Specifically, the robotic mediator is capable of (a) contextually supporting the user (through verbal interaction) in every-day activities, as well as (b) identifying serious emergency situations and issuing appropriate alarms. Two important features of the assistive robot have emerged as very relevant from the user analysis we have presented in this paper: (a) the ability to move robustly in the home environment, e.g., moving smoothly and safely, avoiding obstacles, etc., and (b) the ability to interact naturally with the elderly user. Indeed, these features had already been singled out as important in a preliminary analysis conducted before active development had commenced, and have guided our research throughout the project, particularly with respect to the mobility requirements. These studies will also influence our agenda for future work, particularly concerning aspects related to natural language interaction. Other remarks have emerged from the analysis of tasks that the assistive environment as a whole is able to support. Tasks relative to safety, personal health, and object tracking have been evaluated as critical, and user response to our technological solutions in these areas is extremely positive. This suggests that it is important to foster advancements of assistive technology in these areas. According to our experience, this will require further investment in sensory technology, with a particular emphasis on integrating different types of environmental sensors. Another interesting point which emerges from the evaluation is the relatively low appreciation of “suggestions”. This is an important indication because it poses the question of whether this is due to poor communication capabilities on behalf of the robot, or to an effective lack of interest in these situations. Further analysis will be needed to inspect the possibility of improved added value with more sophisticated natural language interaction. A final comment is orthogonal to the previous ones and concerns the need for multi-disciplinary competences for creating realistic proposals for socially assistive environments. As shown in this paper, the amalgamation of competencies in robotics, artificial intelligence and cognitive psychology has been a driving factor in the development of the RoboCare Domestic Environment. Given the particularly sensitive nature of assistive technology, future challenges in assistive robotics will most likely require an increasing degree of multi-disciplinary research to effectively address the many technical and psychological issues involved. Acknowledgments. This research has been partially supported by MIUR (Italian Ministry of Education, University and Research) under project RoboCare: “A MultiAgent System with Intelligent Fixed and Mobile Robotic Component”, L. 449/97 (http://robocare.istc.cnr.it). References 17 [1] S. Bahadori, A. Cesta, L. Iocchi, G.R. Leone, D. Nardi, F. Pecora, R. Rasconi, and L. Scozzafava. Towards Ambient Intelligence for the Domestic Care of the Elderly. In P. Remagnino, G.L. Foresti, and T. Ellis, editors, Ambient Intelligence: A Novel Paradigm. Springer, 2004. [2] Shahram Bahadori, Luca Iocchi, G. R. Leone, Daniele Nardi, and L. Scozzafava. RealTime People Localization and Tracking Through Fixed Stereo Vision. Lecture Notes on Artificial Intelligence, LNAI 3533:44–54, 2005. [3] P. B. Baltes and M. M. Baltes. Psychological perspectives on successful aging: The model of selecive optimization with compensation. In P. B. Baltes and M. M. Baltes, editors, Successful Aging: Perspectives from the Behavioral Sciences, pages 1–34. Cambridge Univ. Press, New York, 1990. [4] A. Bandura. Self-efficacy: Toward a unifying theory of behavioural change. Psychological Review, 84:191–215, 1977. [5] G. Breakwell. Coping with threatened identity. Methuen, London, 1986. [6] D. Calisi, A. Farinelli, L. Iocchi, and D. Nardi. Information Processing Systems 7, pages 625– Autonomous navigation and exploration in a res632. MIT Press, Cambridge MA, 1995. cue environment. In Proc. of IEEE Interna[15] M.V. Giuliani, M. Scopelliti, and F. Fornara. Eltional Workshop on Safety, Security and Rescue derly people at home: technological help in evRobotics (SSRR), Kobe, Japan, June 2005. eryday activities. In IEEE International Workshop on Robot and Human Interactive Commu[7] A. Cesta, G. Cortellessa, A. Oddi, N. Polinication (ROMAN 2005), pages 365–370, 2005. cella, and A. Susi. A Constraint-Based Architecture for Flexible Support to Activity Scheduling. [16] G. Grisetti, C. Stachniss, and W. Burgard. ImLecture Notes on Artificial Intelligence, LNAI proving grid-based slam with rao-blackwellized 2175:369–381, 2001. particle filters by adaptive proposals and selective resampling. In Proceedings of ICRA-05, [8] A. Cesta, L. Iocchi, G.R. Leone, D. Nardi, pages 2443–2448, 2005. F. Pecora, and R. Rasconi. Robotic, Sensory and Problem Solving Ingredients for the Future [17] H. Z. Haigh, L. M. Kiff, J. Myers, V. GuralHome. Technical report, RoboCare Project nik, K. Kirschbaum, J. Phelps, T. Plocher, , RC-TR-0806-5, 2006. and D. Toms. The independent lifestyle assistant (ILSA): Lessons learned. Technical report, [9] A. Cesta and F. Pecora. Integrating Intelligent Technical Report ACSPO3023, Honeywell LabSystems for Elder Care in RoboCare. In W.C. oratories, Minneapolis, MN, USA, 2003. Mann and A. Helal, editors, Promoting Independence for Older Persons with Disabilities, pages [18] Edwin Hutchins. Cognition in the Wild. MIT 65–73. IOS Press, 2006. Press, 1995. [10] Kerstin Dautenhahn and Iain Werry. Issues of [19] L. Kavraki and J. Latombe. Probabilistic robot-human interaction dynamics in the reharoadmaps for robot path planning. In Practibilitation of children with autism. In Proceedings cal Motion Planning in Robotics: Current Apof the IEEE International Conference on Intelproaches and Future Challenges, pages 33–53. ligent Robots and System, 2002. K.G. and A.P. del Pobil, 1998. [11] Jon Eriksson, Maja J. Mataric’, and Carolee J. [20] M. P. Lawton. Time budgets of older people: A Winstein. Hands-off assistive robotics for postwindow on four lifestyles. Journal of Gerontolstroke arm rehabilitation. In Proc. 9th Int. Conf. ogy, 37:115–123, 1982. on Rehabilitation Robotics (ICORR-05), pages [21] G. J. McAvay, T. E. Seeman, and J. Rodin. A 21–24. IEEE, June 2005. longitudinal study of change in domain-specific [12] D. Feil-Seifer and M. J. Mataric’. Defining soself-efficacy among older adults. Journal of cially assistive robotics. In Proc. 9th Int. Conf. Gerontology, 51:243–253, 1996. on Rehabilitation Robotics (ICORR-05), pages [22] P.J. Modi, W.M. Shen, M. Tambe, and 465–468. IEEE, June 2005. M. Yokoo. Adopt: Asynchronous distributed [13] T. Fong, I. Nourbakhsh, and K. Dautenconstraint optimization with quality guarantees. hahn. A survey of socially interactive robots. Artificial Intelligence, 161:149–180, 2005. Robotics and Autonomous Systems, 42(3-4):143– [23] K. Myers and N. Yorke-Smith. A Cognitive 166, 2003. Framework for Delegation to an Assistive User [14] Bernd Fritzke. A growing neural gas network Agent. In Proceedings of AAAI 2005 Fall Symlearns topologies. In G. Tesauro, D. S. Touretposium on Mixed-Initiative Problem Solving Aszky, and T. K. Leen, editors, Advances in Neural sistants, 2005. 18 [24] Karen Myers. Calo: Building an intelligent per- [31] R. Simmons S. Sabanovic, M.P. Michalowski. sonal assistant. Invited Talk. The Twenty-First Robots in the wild: Observing human-robot soNational Conference on Artificial Intelligence cial interaction outside the lab. In Proceedings and the Eighteenth Innovative Applications of of the International Workshop on Advanced MoArtificial Intelligence Conference (AAAI-06), tion Control, Istanbul, Turkey, March 2006. 2006. [32] M. Scopelliti, M. V. Giuliani, A. M. D’Amico, and F. Fornara. If I had a robot . . . peoples’ [25] F. Pecora, P.J. Modi, and P. Scerri. Rearepresentation of domestic robots. In S. Keates, soning About and Dynamically Posting n-ary P. J. Clarkson, P. M. Langdon, and P. Robinson, Constraints in ADOPT. In Proceedings of 7th editors, Design for a more inclusive world, pages Int. Workshop on Distributed Constraint Rea257–266. Springer-Verlag, London, 2004. soning (DCR-06), at AAMAS’06, 2006. [33] M. Scopelliti, M. V. Giuliani, and F. Fornara. [26] F. Pecora, R. Rasconi, G. Cortellessa, and Robots in a domestic setting: A psychological A. Cesta. User-Oriented Problem Abstracapproach. Universal Access in the Information tions in Scheduling, Customization and Reuse in Society, 4(2):146–155, 2005. Scheduling Software Architectures. Innovations in Systems and Software Engineering, 2(1):1–16, [34] T. Simeon, J. Laumond, and C. Nissoux. Visibility-based probabilistic roadmaps for mo2006. tion planning. In Proc. of IEEE IROS 1999, [27] William Pentney, Ana-Maria Popescu, Shiaokai 1999. Wang, Henry A. Kautz, and Matthai Phili[35] Adriana Tapus, Maja J Mataric’, and Brian pose. Sensor-based understanding of daily life Scassellati. The grand challenges in socially asvia large-scale use of common sense. In Prosistive robotics. IEEE Robotics and Automation ceedings, The Twenty-First National Conference Magazine (RAM) - Special Issue on Grand Chalon Artificial Intelligence and the Eighteenth Inlenges in Robotics, 14(1), March 2007. novative Applications of Artificial Intelligence Conference (AAAI-06), 2006. [36] K. Wada, T. Shibata, T. Saito, and K. Tanie. Analysis of factors that bring mental effects to [28] J. Pineau, M. Montemerlo, M. Pollack, N. Roy, elderly people in robot assisted activity. Proand S. Thrun. Towards Robotic Assisceedings of the IEEE International Conference tants in Nursing Homes: Challenges and Reon Intelligent Robots and System, 2:1152–1157, sults. Robotics and Autonomous Systems, 42(3– 2002. 4):271–281, 2003. [37] S. L. Willis. Everyday cognitive competence in elderly persons: Conceptual issues and empirical [29] Martha E. Pollack, Laura Brown, Dirk Colfindings. The Gerontologist, 36:595–601, 1996. bry, Colleen E. McCarthy, Cheryl Orosz, Bart Peintner, Sailesh Ramakrishnan, and Ioannis [38] S. Wilmarth, N. Amato, and P. Stiller. Maprm: Tsamardinos. Autominder: an Intelligent CogA probabilistic roadmap planner with sampling nitive Orthotic System for People with Memory on the medial axis of the free space. TechniImpairment. Robotics and Autonomous Systems, cal report, Department of Computer Science, 44(3–4):273–282, 2003. Texas A&M University, College Station, TX, nov. 1998. [30] M.E. Pollack. Intelligent Technology for an Aging Population:The Use of AI to Assist El- [39] A. V. Wister. Environmental adaptation by persons on their later life. Research on Aging, ders with Cognitive Impairment. AI Magazine, 11:267–291, 1989. 26(2):9–24, 2005. 19