Using proprioceptive sensors for categorizing interactions

Transcription

1 Using proprioceptive sensors for categorizing interactions [Extended Abstract] T Salter, F Michaud and D Létourneau Université de Sherbrooke Sherbrooke Quebec, Canada t.salter f.michaud ABSTRACT Increasingly researchers are looking outside of normal communication channels (such as video and audio) to provide additional forms of communication or interaction between a human and a robot, or a robot and its environment. Amongst the new channels being investigated are infrared, proprioceptive and temperature sensors to detect touch. Our work aims at developing a system that can detect natural touch or interaction coming from children playing with a robot, and adapt to this interaction. This paper reports trials done using Roball, a spherical mobile robot, demonstrating how sensory data patterns can be identified in human-robot interaction, and exploited for achieving behavioral adaptation. The experimental methodology used for these trials is reported, which validated the hypothesis that human interaction can not only be perceived from proprioceptive sensors on-board a robotic platform, but that this perception has the ability to lead to adaptation. General Terms Algorithms, Performance, Design, Experimentation, Human Factors. Keywords Human-Robot Interaction (HRI), Adaptive Mobile Robots, Sensor Evaluation, Categorizing Interaction. (Produces the permission block, copyright information and page numbering) A full version of this paper is available as Author s Guide to Preparing ACM SIG Proceedings Using L A TEX2 ɛ and BibTeX at Also affiliated with University of Hertfordshire D.C. Lee and I.P. Werry University of Hertfordshire Hatfield Hertfordshire, England d.c.lee 1. INTRODUCTION Touch is an important form of communication or interaction between robots and humans [16], [17] or between robots and their environment [6]. While video and audio are typically used for communication and interaction between humans and robots, other sensors such as contact, infrared, proprioceptive and temperature sensors can provide additional means of communication related to touch [6] and [13]. Our main aim is to develop a system where children can interact with a robot naturally and the robot can adapt to this natural interaction. How children interact with and perceive robots is itself becoming a highly studied area [3], [14], [23], [5], [22]. People working with children or in therapy are beginning to recognize that natural touch is an important form of interaction or communication with a robot [3], [21], [16], [17], [2]. In an effort to register touch or communication, some robotic systems utilize buttons that must be pushed by a person [4], [1], [19]. Salter et al. [16], [17], [15] showed that infrared sensors on-board a mobile robot, typically exploited for navigation purposes, can also be used to record interactions or natural touch coming from children playing with a mobile robot. The research demonstrates that it is even possible to detect personality traits (e.g., boisterous, or cautious) of a child interacting with a wheeled robot, simply from the analysis of infrared sensor data. At MIT [20], a robotic teddy bear named Huggable is being designed with full-body sensate skin and smooth, quiet voice coil actuators that are able to relate to people through touch. Huggable features a series of temperature, electric field and force sensors which it uses to sense the interactions that people have with it [12]. At NASA, Lumelsky [8] is working on developing a sensitive skin that could be used to cover a robot. The skin would include more than 1,000 infrared sensors that would detect an object [7]. In their related work, Saraf and Maheshwari [9] claim that their device can give a robot tactile sensitivity equivalent to that of human fingers, and one early use might be in minimally invasive surgery. The system they have developed is based on alternating layers of gold and semi-conducting cadmiumsulfur nanoparticles separated by nonconducting, or dielectric, films. They hope to coat a robot hand with this film. Another move away from the typical use of microswitches for detecting touch is realized in the robot seal named Paro, which was developed for

2 Figure 2: Both front or back and side view. Figure 1: Roball, an autonomous rolling robot. robot assisted activity in hospitals or homes for the elderly. In this system, physical contact with the robot, for example touching, is recognized by a system based on balloons [21]. Our interests lie in studying how robots proprioceptive sensors (which can be used for navigation, control or other purposes) can be exploited as a form of communication and as way to capture natural touch or interaction between a mobile robot and children. First, a series of trials were conducted using Roball (see Figure 1) in laboratory conditions without children and then in real life settings with children. Data analysis of accelerometers and tilt sensors established that it is possible to detect play patterns with accelerometers and tilt sensors. Second, another series of trials were designed to follow on from and expand upon this work, pursuing the goal of adapting the robots behavioral response to the stimulation or interaction it is receiving from children playing with it. The trials were conducted to test whether heuristics identified in the previous trials could be utilized to enable adaptation of Roball to human interaction. Again trials were conducted using Roball (see Figure 1) both in the laboratory without children and in a real life setting with children. Our findings are presented in this paper. 2. ROBALL Shown in Figure 1, Roball is 6 inches in diameter and weighs about 4 pounds [10], [11]. It consists of a plastic sphere (a hamster exercise ball) constructed from two halves that are attached to each other. The plastic sphere is used to house the fragile electronics (sensors, actuators, processing elements), thus making it robust and ideal for interaction with children. The fact that Roball is spherical encourages a wide range of play situations. Movement is achieved through a combination of two propulsion motors that are used to propel the shell of Roball, and a counterweight that is used to control direction by moving the center of gravity towards Figure 3: Roball s accelerometers, centered above the steering motor. the required direction, see Figure 2. The propulsion motors are attached to the shell wall. Rotation of these motors causes Roball to move forward or backwards. The steering motor moves the counterweight from one side to the other in order to move the center of gravity away from the center of the sphere. Roballs first prototype used binary mercury switches to detect and control lateral and longitudinal inclinations of the internal plateau. To provide a wider set of perceptual states, new sensors were installed on the platform. The version of Roball used in our work, running with a PIC18F458 microcontroller with 32 kbytes of internal memory, 1.5 kbytes of RAM and operating at 40 MHz (10 MIPS), has the following special features: Accelerometers - Roball has three accelerometers, one for each axis (X, Y and Z). Analog ADXL311 miniature accelerometers are used to measure Roballs acceleration in any of the three axes directions in the range 0 to 2g (see Figure 3). Tilt sensors - There are three tilt sensors, one for left tilt, one for right and one for forward/backward tilt. Sharp GP1S036HEZ miniature photointerrupters are used to detect the tilt direction by sensing the movement of a small ball. The tilt sensors are positioned on Roballs printed circuit board. Two tilt sensors are placed symmetrically on left/right axis (the axis cor-

3 Figure 5: Examples of the experiments in the laboratory. The robot executed a wandering and a simple obstacle avoidance behaviour. Figure 4: Roball s tilt sensors: (a) placement; (b) possible values with L = Left, R = Right, B = Back and F = Front; (c) left tilt being registered. responding to the line between Roballs two propulsion motors). This configuration allows the detection of either left or right tilt with both sensors giving the same value, and also allows detection of rotation with readings from the sensors giving opposite left/right tilt values due to centrifugal acceleration. For example, if the ball is tilted to the left, both the right and the left tilt sensors give a reading of (L). If the ball is tilted to the right, both the right and left tilt sensor gives a reading of (R). Finally, if the ball is spinning the right and left tilt sensors give opposite readings: the right sensor gives a reading of (R) and the left sensor gives a reading of (L) (see Figure 4). 3. EXPERIMENTAL SETTINGS In both sets of trials (First and Second), the robot was programmed to execute two simple behaviors: wandering and obstacle avoidance. These behaviors were carried out for the duration of each trial. Roball s function was to act as a mobile, moving toy. Four small wooden walls enclose the experimental arena, which creates a pen as shown in Figures 5 and 6. The pen is approximately 2.5 m by 2 m. Every experiment was video taped for verification of sensor readings. Sensor readings were recorded on-board the robot every 0.10 second. After the preprogrammed duration of the trial the robot stops by itself, ending the trial. All of the laboratory experiments took place at the Université de Sherbrooke, without children present. In the real life settings, the children were typically developing boys aged between 5 and 7 years. All the children were treated the same. None of the children had been exposed to robotic devices other than those found in toy stores. When the trial begins they are asked to step inside the pen and to play Figure 6: The children playing with Roball in a home setting. Roball is programmed with adaptive behaviours. with Roball. The experimenter attempts to have as little effect on the children as possible. Such as giving the same set of instructions to all the children, also not engaging in conversation with children and not prompting the children to play with the robot once the trial has commenced. 4. FIRST TRIALS: OBTAINING HEURIS- TICS The first set of trials involved: a series of laboratory experiments then a series of trials held in real life settings, namely, a play group and a school. 4.1 Laboratory Experiments These were used to investigate whether measurements from two different types of proprioceptive sensors could record things such as jolts to the robot, the robot receiving general interaction, the robot being carried or the robot being spun. The experiments were broken down into seven environmental conditions: 1. Alone (i) Roball wandering in the laboratory. 2. Alone (ii) Roball wandering in the experimental pen. 3. Light Boxes Roball wandering in the pen with boxes present. 4. Heavy Boxes - Roball wandering in the pen with weighted boxes present. 5. Carrying Experimenter walking whilst carrying Roball.

4 Figure 7: The 3 different axes reading for each of the environmental conditions. Figure 8: Averaged differences between the X and Z accelerometer readings. 6. Interaction Experimenter simulating interaction e.g. kicking, pushing, banging Roball. 7. Spinning Experimenter spinning Roball. Data analysis was performed after the experiments had been conducted. Analysis was conducted by using simple calculations on the data collected from the accelerometers and tilt sensor during these seven different environmental conditions. The data comes from an average of the 3 experiments of 5 minutes each, repeated for each of the seven environmental conditions. Data is being recorded at 10 Hz Accelerometer Results Figure 7 shows that during the seven different environmental, each of the accelerometers produces a different (average) signature. Although there are similarities between the signatures, each is unique in some way. Figure 8 illustrates that the difference between the X accelerometer and the Z accelerometer (X Z) also produced very interesting results that could be used to discriminate the robots interaction status. For instance, only in condition (6) Carrying do we see a negative X-Z difference; also, condition (1) Alone (i) shows the largest difference between these two axes. Line graphs produced of the trials clearly show a difference in the way human contact with robot is registered by accelerometers to what these sensors register when the robot is alone (see Figures 9 and 10). These graphs show a marked difference in sensor readings and identify that there definitely is the possibility of using these readings to correctly classify human interaction and thus have the ability to adapt to this information. When the robot has experienced interaction we see all three axes readings show jagged readings that constantly cross with each other. When the robot is, alone, we see large gaps between each of the axis Tilt Sensor Results Data recorded from the tilt sensors readings during each of the When the robot is spinning, two tilt sensors should give different values as described in Section 2 (see Figure Figure 9: The erratic X, Y and Z axis sensor data from the accelerometers when Roball is in experimental condition (6) Interaction. Figure 10: Accelerometer data when Roball is in experimental condition (1) Alone(i).

5 Figure 11: Tilt sensor readings for the seven environmental conditions. 4). From the results shown in Figure 11, we can see that condition (7) Spinning, as expected, produces the highest results. 4.2 Real Life Settings Two real life setting were used to confirm that the data found under laboratory conditions would also be found in the real world. A play group and a school setting were used. In total eight boys participated in the trial, each playing with Roball at least twice. Trials were initially conducted for 5 minutes but this seemed to long for the children to hold their attention span and so the length of the trials were shortened to 4 minutes. Similar results were obtained from the analysis of accelerometers, tilt sensors and line graphs from the real life settings to those found in the lab. As with the trials conducted in the laboratory we found that Cautious children that did not play with Roball very much receive the largest difference between the X - Z axes, see Figure 8 for the results from the laboratory. We get the highest number of different tilt sensor readings for active children that spin Roball, see Figure 11 for the results from the laboratory. When a child plays with Roball we observe that interaction can be seen as jagged lines. When the child did not play with Roball we see gaps between the different axes (see Figure 12). 4.3 FIRST TRIAL: Conclusions This first set of trials showed that it is possible to detect different environmental conditions through the analysis of proprioceptive sensors (accelerometers and tilt) [18]. Overall, these analyses indicate that different environmental conditions, which can be associated with forms of interaction, can be detected through the analysis of proprioceptive sensor data. The detection that the robot is being carried is the easiest, followed by the detection that the robot is being spun. The detection of general interaction with a person is not so easy, but still possible, however it is easier to detect when the robot is alone, or not receiving any type of interaction from a person. To categorize between these states, Figure 12: Accelerometer readings when a child was interacting with Roball in a real life setting. the sensor reading space can be zoned into regions. The objective is to achieve this categorization using an on-board feature detection algorithm, in real-time, and to adapt the robots behavior to the interaction experienced. 5. HEURISTICS FOR BEHAVIORAL ADAP- TATION It was found in the first trial that it is possible to detect different environmental conditions. More specifically, related to interaction with people, accelerometer and tilt readings that can be classified into zones which detect four modes of interaction.: ALONE, GENERAL INTERACTION, CAR- RYING and SPINNING. Another condition, named NO CON- DITION, is necessary for situations not part of the other four. By detecting these states, the robots behavior could be changed or adapted to respond in a particular fashion to interaction with children. The algorithm developed uses the following five rules which are specific instantiations of the heuristics derived from the analysis presented in Section 4 and in [18]. A Being Alone. If the average difference between the X and Z accelerometer readings is above 0.05, set current mode to ALONE. B Receiving General Interaction. If the average difference between the X and Z accelerometer readings is below 0.03 and above zero, set current condition to GENERAL INTERACTION. C Being Carried. If the average difference between the X and Z accelerometer readings is negative, set current condition to CARRYING. D Being Spun. If the tilt sensors show different readings (see Figure 11), set condition to SPINNING. Another way to detect spinning is if the average reading for the Z axis is positive and coupled with an average Y axis reading of above 0.05 (see Figure 7). E No Condition. If the sensor readings do not fall into one of the above categories, set the condition to NO CONDITION.

6 Different from the first trial, the algorithm uses a temporal window of 4 seconds to calculate an average of the sensor readings and thus derive which condition it believes the robot is currently experiencing. This window is moved forward in time by 0.10 sec increments. After examining the heuristics it was decided to develop an algorithm that detected conditions based on the ease with which it was believed they were detectable i.e. checking for conditions that are easy to classify first. The final flow of the algorithm was determined by various tests conducted in the laboratory. Tests were conducted that checked the ability of the algorithm to detect the differing conditions until the optimal flow for the algorithm was achieved. The algorithm is designed to first attempt to detect the condition SPINNING, by looking at the difference in the tilt sensor readings. If this condition is found to be true, no analysis of the accelerometers is carried out. However, if this is found to be false, the analysis of the accelerometers is then performed. For this analysis, the condition CAR- RYING is first checked by looking for a negative (X Z) average. Next, the condition GENERAL INTERACTION is looked for by seeing if the Z axis is below zero. Then, the condition SPINNING is analysed by checking for a positive Z axis coupled with an average Y axis reading of above the condition ALONE is then identified if the X - Z average is above Next, the condition GENERAL IN- TERACTION is classified if the X - Z average is below 0.03 but above zero. If, after all this, the current average within the four second window does not fall into any of the categorises, NO CONDITION is set as the output. The resulting pseudo-code of the algorithm is: Being Carried The experimenter walked whilst carrying Roball for the duration of the experiment. Receiving General Interaction The experimenter simulated interaction from a child, for example, pushing, banging and getting in the way of the robot. Being Spun The experimenter purposely spun the robot for the duration of the experiment. Three separate trials were conducted for each of the four conditions and each individual trial lasted again for a duration of 4 minutes. Thus, in total, 12 experiments were carried out, lasting a total of 48 minutes. This resulted in 2360 interaction classifications (10 per second for 240 seconds less 40 for the four seconds in which the algorithm is initializing i ts temporal average window). The adaptation algorithm was implemented on-board Roball, and the identified states were recorded as interaction occurred. Table 1 presents the observed results of the identified states (A = ALONE, B = GENERAL INTERACTION, C = CARRY- ING, D = SPINNING, E = NO CONDITION) in relation to the four modes of interaction Algorithm Results The results represent the percentage of time that the state was identified during the trial. The objective is to maximize valid identification and minimize false detection. As can be seen from the leading diagonal of Table 1, the robot can identify the following with reasonable accuracy: IF tilt sensors are different output SPINNING ELSE %Check accelerometers IF x-z average < 0 output CARRYING ELSE IF z average < 0 output GENERAL INTERACTION ELSE IF z average > 0 && y average > 0.05 output SPINNING ELSE IF x-z average > 0.05 output ALONE ELSE IF x-z average > 0 && x-z average < 0.03 output GENERAL INTERACTION ELSE output NO CONDITION END IF END IF 6. SECOND TRIALS: BEHAVIORAL ADAP- TATION A second series of trials were conducted that mimicked those carried out in the first set of trials. Again trials were conducted in the laboratory and a real life setting. 6.1 Laboratory Experiments The experiments were broken down into the four modes of interaction listed below: Being Alone - Roball wandered in the pen by itself, no objects or humans present. Being Alone (97%) Being Carried (92%) Being Spun (77%) However, identifying Receiving General Interaction (10%) is revealed to be more difficult. A probable cause for this is that at times the robot is in fact SPINNING or experiencing ALONE during the General Interaction trials. Such conditions would therefore be identified under the corresponding categories, (D) SPINNING 45% and (A) ALONE 19% of the time. Therefore, adding the results for conditions (A) ALONE, (B) GENERAL IN- TERACTION and (D) SPINNING, a total of 74% classifications were correctly identified by the algorithm during the GENERAL INTERACTION experiment. It should be noted that the experimenters simulation of general interaction was fairly vigorous. For example, the experimenter pushed and kicked the robot with quite some force, which did cause the robot to spin. This is not always the case with children, as can be seen from the line graph of a child s interaction, see Figure 12. In general, from observations of the first set of trials, children seem to interact with the robot in a more punctuated manner, so that a general interaction is rather rare. For example, they may push the

7 robot then wait, or spin the robot but not then straight away push and kick the robot. Therefore, it is believed that recording the varying interaction may possibly be easier in a real life setting than in the laboratory. The misclassification of the condition CARRYING whilst receiving general interaction as shown in Table 1 requires further discussion. It was discovered that when the robot hits the wall of the pen, it actually records the same readings as being carried. This is because as the robot hits the pen wall, it slightly rolls up the wall and this causes similar accelerometer readings as when the robot is picked up. This therefore helps to explain why when the robot is Receiving General Interaction, the condition CARRYING was registered 19% of the time and similarly when the robot was Being Spun, the condition CARRYING was registered 17%. During both trials the robot did roll up the wall of the pen. Where as, when the robot was Alone it did not roll up the wall thus we see the condition CARRYING recorded 0.5%. The problem of the robot rolling up the wall was not considered significant as the pen was only being used specifically in this research work. In other real life settings the robot would not be so confined and therefore this phenomena would not happen so often. *** 6.2 Adding Adaptation and Testing in a Real Life Setting Three adaptive behaviors were added to the robot: two behaviors involved vocals and one behavior involved motion coupled with vocals. 1. When the robot classifies its sensor readings as SPIN- NING the robot produces the sound: weeeeeeeeeeeeee. 2. When the robot classifies its sensor readings as CAR- RYING it stops all motion and says put me down. 3. When the robot classifies its sensor readings as ALONE it says play with me. Both the sounds and the movement response are repeated until the condition that initiated the behavior changes. Audible responses are easily perceived by an external observer, which facilitates post-experiment analysis compared to a change in the behavioral control of the robot. In this work it was decided to begin simply with limited adaptation for clear and precise evaluation of the algorithms performance in the child-robot trials. The experiments in a real life setting took place at a house, as shown in Figure 6. All the children were known to the experimenter in a social context. Different to the first trial, the children were told two things about the robot: 1) the robot could be spun; 2) they could pick the robot up but they must not then drop it and that they should put the robot gently down on the floor. Each trial with the children lasted for four minutes. From the preliminary results observed with children interacting with the robot in trial, we clearly observe that the robot did react to the children and that when the robot did react there was an increased level of involvement by the children. However, it was noticed that at times the robot did not react correctly. One clear case is the identification of Carrying when it hit the wall of the pen, causing the robot to stop all motion and say Put me down. Interestingly, as a side effect, this caused an even higher level of engagement and interaction from the children: e.g., having the child looking at the experimenter and saying it is asking me to put it down and then proceeding to aid the robot by moving it so that it could progress on its way. Overall, the response of the children was very encouraging. It shows that in principle that is possible to adapt the behavior of a robot using sensor readings. Evident was that there was a increase in the level of the childrens engagement and interaction in second trial compared to that of the first trial. It seems that this was due to a more complex behavior. Increasing levels of engagement and interaction with children, which was the ultimate goal of the robot. A further and detailed analysis of the child-robot study needs to done. Also we need to conduct further trials with a larger sample size of children and we need to fully investigate the reasons for any incorrect categorization of the algorithm, which ultimately leads to incorrect reaction of the robot. 7. CONCLUSIONS This work demonstrates that proprioceptive sensors are capable of detecting and recording human interaction. Through the analysis of accelerometers and tilt sensors (proprioceptive sensors) on the Roball s platform, it is found that that is possible to detect the robot being carried, being spun, being alone and also interacting with a person. The analysis required is simple, work in real-time on a small embedded microcontroller, and does not require complex processes. Carrying out preliminary trials in the laboratory without children helped in providing base line and bench marks readings for successive child-robot trials. Further work will include more detailed analysis of the adaptation algorithm used with Roball interacting with children. Also planned is the use of the adaptation algorithm to change the robots navigational behavior, and see the effects on engagement and interaction with children. 8. ACKNOWLEDGMENTS F. Michaud holds the Canada Research Chair (CRC) in Mobile Robotics and Autonomous Intelligent Systems. This work is funded by the CRC and the Canadian Foundation for Innovation. 9. REFERENCES [1] A. Billard. Robota: Clever toy and educational tool. Robotics and Autonomous Systems, 42: , [2] A. Duquette, H. Mercier, and F. Michaud. Investigating the use of a mobile robotic toy as an imitation agent for children with autism. In International Conference on Epigenetic Robotics, Paris, France, [3] T. Ito and N. P. R. Center. How children perceive robots. univ/05/univ e05.html, last accessed 06/10/04, 2003.

8 Table 1: The rows of the table correspond to the four tested environmental conditions. The columns of the table correspond to the five categories derived from the sensor readings, A = ALONE, B = INTERACTION, C = CARRYING, D = SPINNING, E = NO CONDITION. The percentage of correctly classified interaction is therefore shown in the leading diagonal. For example, the robot being alone is correctly classified 97% of time. A B C D E Being Alone 97% 0.5% 0.5% 1% 1% Receiving General Interaction 19% 10% 19% 45% 7% Being Carried 0.5% 3% 92% 3% 1.5% Being Spun 2% 3% 17% 77% 1% [4] B. Jensen, G. Froidevaux, X. Greppin, A. Lorotte, L. Mayor, M. Meisser, G. Ramel, and R. Siegwart. The interactive autonomous mobile system robox. In IROS 2002, IEEE/RSJ International Conference on Intelligent Robots and Systems, pages IEEE Press, [5] T. Kanda, T. Hirano, D. Eaton, and H. Ishiguro. Interactive robots as social partners and peer tutors for children: A field trial. Human-Computer Interaction, 19:61 84, [6] O. Kerpa, K. Weiss, and H. Worn. Development of a flexible tactile sensor system for a humanoid robot. In IEEE International conference on Intelligent Robots and Systems (IROS 2003). IEEE Press, [7] V. Lumelsky. NASA - vladskin.html. Last accessed [8] V. Lumelsky, M. Shur, and S. Wagner. Sensitive skin. IEEE Sensors Journal, 1, [9] V. Maheshwari and R. Saraf. High-resolution thin-film device to sense texture by touch. Science, pages , June [10] F. Michaud and S. Caron. Roball, the rolling robot. Autonomous Robots, 12(2): , [11] F. Michaud, J.-F. Laplante, H. Larouche, A. Duquette, S. Caron, D. Létourneau, and P. Masson. Autonomous spherical mobile robot for child-development studies. Systems, Man, and Cybernetics, 35: , [12] MIT. ResearchPubWeb.pl?ID=53, Last accessed 24/08/06. [13] T. Miyashita, T. Tajika, H. Ishigurio, K. Kogrue, and N. Hagita. Haptic communication between humans and robots. In 12th International Symposium of Robotics Research, San Francisco, CA, USA, [14] B. Robins, K. Dautenhahn, R. te Boekhorst, and A. Billard. Effects of repeated exposure of a humanoid robot on children with autism. Universal Access and Assistive Technology (CWUAAT), pages , [15] T. Salter and K. Dautenhahn. Guidelines for robot-human environments in therapy. In IEEE Ro-man 2004, 13th IEEE International Workshop on Robot and Human Interactive Communication, pages 41 46, Kurashiki, Okayama, Japan, IEEE Press. [16] T. Salter, K. Dautenhahn, and R. te Boekhorst. Learning about natural human-robot interaction. Robotics and Autonomous Systems, 54(2): , [17] T. Salter, K. Dautenhahn, and R. te Boekhorst. Robots moving out of the laboratory - detecting interaction levels and human contact in noisy school environments. In IEEE Ro-man 2004, 13th IEEE International Workshop on Robot and Human Interactive Communication, pages , Kurashiki, Okayama, Japan, IEEE Press. [18] T. Salter, F. Michaud, K. Dautenhahn, D. Létourneau, and S. Caron. Recognizing interaction from a robot s perspective. In RO-MAN 05, 14th IEEE International Workshop on Robot and Human Interactive Communication, Nashville, TN, USA, [19] Sony. lasted accessed 06/10/04. [20] W. D. Stiehl, J. Lieberman, C. Breazeal, L. Basel, L. Lalla, and M. Wolf. The design of the huggable: A therapeutic robotic companion for relational, affective touch. In AAAI Fall Symposium on Caring Machines: AI in Eldercare, Washington, D.C., [21] D. Wada, T. Shibata, T. Saito, and K. Tanie. Robot assisted activity for elderly people and nurses at a day service center. In IEEE International Conference on Robotics and Automation, pages , Washington, DC, [22] T. Watanabe, R. Danbara, and M. Okubo. Interactor: Speech-driven embodied interactive actor. In IEEE Ro-man 2002, 11th International Workshop on Robot and Human Interactive Communication, pages , Berlin, Germany, IEEE Press. [23] S. Woods, K. Dautenhahn, and J. Schulz. Child and adults perspectives on robot appearance. In AISB 05 Symposium on Robot Companions: Hard Problems and Open Challenges in Robot-Human Interaction, Hertfordshire, England, UK, 2005.