Extracting Multimodal Dynamics of Objects Using RNNPB

Transcription

1 Paper: Tetsuya Ogata Λ, Hayato Ohba Λ, Jun Tani ΛΛ, Kazunori Komatani Λ, and Hiroshi G. Okuno Λ Λ Graduate School of Informatics, Kyoto University, Kyoto, Japan fogata, hayato, komatani, ΛΛ Brain Science Institute, RIKEN, Saitama, Japan [Received February 4, 2005; accepted April 30, 2005] Dynamic features play an important role in recognizing objects that have similar static features in color or shape. This paper focuses on active sensing that exploits the dynamic feature of an object. An extended version of the robot, Robovie-IIs, uses its arms to move an object and determine its dynamic features. At issue is how to extract symbols from different temporal states of the object. We use a recurrent neural network with parametric bias (RNNPB) that generates self-organized nodes in parametric bias space. We trained an RNNPB with 42 neurons using data on sounds, trajectories, and tactile sensors generated while the robot was moving or hitting an object with its arm. Clusters of 20 types of objects were selforganized. Experiments with unknown (untrained) objects showed that our proposal configured them appropriately in PB space, demonstrating its generalization. Keywords: active sensing, humanoid robot, recurrent neural network 1. Introduction We are developing techniques to enable robots to manipulate tools designed for human use. Conventional robots manipulate specific tools designed for robot hands, and it remains difficult for mechanical systems to handle dynamics of objects and generate adaptive behavior by learning a dynamic environment. In tool manipulation, for example, robots identify objects but do not yet predict an object s dynamics. Tool manipulation is crucial to recognizing objects recognition and several studies have been done on active sensing [1 5]. Noda et al. reported using a humanoid robot, Wamoeba-2Ri, which grasps objects with its hand and recognizes them by integrating multiple sensory data on size, weight, and color imaging [2]. This study used a three-layered self-organizing map (SOM) [6], which deals only with static features, and required over a thousand neurons for processing multimodal sensory data, making it difficult for the robot to apply recognition results in motion planning. Arsenio et al. focused on rhythmic motion as object dynamics and merged audiovisual sensory data to recognize objects using the humanoid robot Cog [3]. This study showed that cross-modal dynamics are essential for object recognition and manipulation, but targeted only rhythmic motion generated by human operators, rather than the robot. This did not enable the robot to plan more generalized tool manipulation. Fukano et al. presented an example in which a multifingered hand grasps the target object, observes the resulting motion through proprioception, and detects motion constraints [4]. This study yielded interesting results on active sensing with stochastic techniques, and the unknown objects were bottles of different sizes. The common problem of the above studies is that their target was recognition of fewer than 10 objects designed and/or selected for the robot. We propose novel active-sensing using object dynamics a recurrent neural net (RNN) trained using multimodal sensory data generated while a robot is moving or colliding with objects. The RNN enables robots to use dynamic features for different object recognition and motion prediction. Our proposal is generalizable enough to configure unknown (untrained) objects appropriately. Section 2 introduces the recurrent neural network model for learning. Section 3 details actual design of active sensing, e.g., motion design, target objects, sensors, and the neural network configuration. Section 4 discusses experiments and results of our proposal. Section 5 discusses the characteristics of our proposal and compares them to those of conventional recognition. Section 6 concludes this paper and reviews projected work on motion generation. 2. Learning Algorithm This section details how robots deal with dynamic features of sensory information during active sensing. Statistical techniques represented by the hidden Markov model (HMM) process time-sequence data efficiently, but require huge amounts of data for learning. This becomes a problem for real robots conducting experiments for collecting data due to the durability problem. The HMM deals only with known objects, which could be fatal in adaptability to a real dynamic environment. We used a deterministic method represented by an artificial neu- Journal of Robotics and Mechatronics Vol.17 No.6,

2 Ogata, T. et al. Fig. 2. External ears and hand. Fig. 1. RNNPB network configuration. ral net (ANN) to solve this problem. that the RNN selforganizes (acquires) contextual information [7]. We use the forwarding forward (FF) model proposed by Tani [8], also called the RNN with parametric bias (RN- NPB) model. It articulates complex motion sequences into motion units, which are encoded as limit-cycling or fixed-point dynamics of the RNN. We previously reported a study of human-robot interaction based on quasisymbols acquired by the RNNBP [9] RNNPB Model The RNNPB model has the same architecture as the conventional Jordan RNN model [10] except for parametric bias (PB) nodes in the input layer. Unlike other input nodes, PB nodes take a constant value throughout each time sequence and are used to implement mapping between fixed length and time sequences. The network configuration of the RNNPB model is shown in Fig.1. As with the Jordan RNN model, the RNNPB model learns data sequences supervisedly. The difference is that in the RNNPB model, values encoding sequences are self-organized in PB nodes during learning. Common structural properties of training data sequences are acquired as connection weights by using the back propagation through time (BPTT) algorithm [11], also used in the conventional RNN. The specific properties of each individual time sequence are simultaneously encoded as PB values, so the RNNPB model self-organizes mapping between PB values and time sequences PB Vector Learning The learning algorithm for PB vectors is a variant of the BPTT algorithm. The step length of a training sequence is denoted by l. For each sensory-motor output, backpropagated errors for PB nodes are accumulated and used to update PB values. Update equations for the ith unit of the PB at the t in the sequence are as follows: ρ t = k bp t+l=2 δt bp + k nb (ρ t+1 2ρ t + ρ t 1 ). (1) t l=2 p t = sigmoid(ρ t =ζ ) (2) In Eq.(1), the δ force for updating internal values of PB p t is obtained by summing two terms. The first term represents delta error, δt bp, back propagated from output nodes to PB nodes: it is integrated over the period from the t l=2 tothet + l=2 steps. Integrating delta error prevents local fluctuations in output errors from significantly affecting temporal PB values. The second term is a lowpass filter that inhibits frequent rapid changes in PB values. k bp, k nb, and ε are coefficients. Current PB values are obtained from sigmoidal outputs of internal values. After learning sequences, the RNNPB model generates a sequence from corresponding PB values. The RNNPB model can be used for recognition and sequence generation. For a given sequence, the corresponding PB value is obtained by using update rules for PB values, without updating connection weights. This inverse operation for generation is regarded as recognition. The RNNPB model acquires relational structures among training sequences in PB space through learning. This generation enables the RNNPB model to generate and recognize unseen sequences without additional learning. By learning several cyclic time sequences for different frequencies, for example, it generates novel time sequences of intermediate frequencies. In the RNNPB, step length l should be determined by dynamic properties of target sequences such as the learning coefficient, but acquired PB patterns were almost the same for a wide range of l. Adjustment is not strict in practical sequence learning. 3. Active Sensing by Moving Objects 3.1. Adding New Functions to Robovie-IIs We refined the humanoid robot Robovie-IIs [12] as a platform for our experiments. Robovie-IIs itself is refined from Robovie-II developed at ATR [13]. The original Robovie-II has three degrees of freedom (DOF) at the neck and four DOF at each arm. It uses two CCD cameras on the head. Tactile sensors of soft rubber silicon cover Robobie-IIs and distinguish among three types of contact collision, rubbing, and touch by detecting changes in pressure. We added functions to Robovie-IIs (Fig.2), i.e., two external ears on the head and two 1-DOF hands on the arms for the experiment on active sensing. 682 Journal of Robotics and Mechatronics Vol.17 No.6, 2005

3 Fig. 5. Object recognition system configuration. Fig. 3. Experiment in active sensing RNNPB Configuration and Learning The following sensory data was normalized ([0-1]) and synchronized (9 frame/s) between different modalities for use by the RNNPB model: 1) Audio Information (5): An audio signal was detected by microphones in the robot s ears (48kHz). Five signal features were extracted using a Mel Filter Bank. 2) Visual Information (4): The center (x; y) and color (R;B) were detected by a CCD camera with a resolution of pixels (30 frame/s). Fig. 4. Recognition targets Motion of Active Sensing and Target Objects Perception using only static features such visual images is not sufficient for distinguishing among objects having similar sizes, shapes, and colors. Such a recognition framework is not applicable to dynamic motion planning. Perception must be designed for sensory-motor coordination [14]. We focused on active sensing motion in which a robot moves or hits an object on a table with its arm. Infants often touch and hit unknown objects in front of them, thereby acquiring skill in manipulating. The movement of a moving object enables the object s dynamic features to be exploited, e.g., tactile pressure to move it, actual trajectory, and sound patterns generated by collision with the table. Figure 3 shows an experiment in which Robovie touched and moved an object by rotating a shoulder motor (roll axis) with constant velocity (60 ffi =s). As the robot moved an object, it collected sound, object trajectory, and touch pressure using its microphones, cameras, and tactile sensors. Figure 4 shows the 20 types of objects used as recognition targets rubber ball, plastic ball, ceramic cup, 2 types of plastic cup, glass, can, moneybox, stuffed doll, Rubik cube, toy car, funnel, pen tray, scrub brush, soft brush, water dumbbell, and shampoo container. The water dumbbell and shampoo container were full or empty two conditions not distinguishable statically as, for example, in visual images. 3) Tactile Information (1): Input voltage from tactile skin sensors was used (4.3Hz). The system configuration is shown in Fig.5. The RN- NPB works as a prediction system whose input is current sensory data s(t) and whose output is the next sensory state s(t + 1): It consists of 42 neurons 10 in the input layer, 20 in the middle layer, 10 in the context layer, and 2asPB. The training sequence of the RNNPB was segmented when changes in all sensory input were less than the threshold. In experiments, sensory sequence lengths L s were 15 to 40 steps. Our goal was to acquire specific parameters corresponding to each object for recognition and motion generation. To fix parameters during sensing, Eq.(1) was simplified in our RNNPB model training as follows: ρ t = k bp t+l=2 δt bp (3) t l=2 Equation (2), which normalizes parameters, was not used in our experiments for analyzing the acquired PB. 4. Experiments and Results 4.1. Self-Organization of PB Space and Modality Differences We conducted an experiment using the 20 objects in the previous section to confirm clustering by our proposal using dynamic features of objects. Robovie moved objects five times (20 5 = 100 sequence data), and the RN- NPB was trained by collecting 100,000 samples, which Journal of Robotics and Mechatronics Vol.17 No.6,

4 Ogata, T. et al. (a) Example of a glass (a) PB space acquired by tactile sensor (b) Example of a scrub brush (b) PB space acquired by audio signal Fig. 6. Sensor flow and prediction output of the RNNPB. required approx 1 hour using a PC with a Pentium IV processor (2.8GHz). Figure 6 shows the sequence of tactile pressure, object position (x coordinate), and sound, when Robovie moved (a) a glass and (b) a scrub brush. Black lines indicate RNNPB input (real value) and gray lines RNNPB output (prediction). We confirmed that the RNNPB predicts each sequence well. Average prediction error is less than 1.5%. Figure 7 shows PB space acquired by each sensor modality. Two parameters in the RNNPB before normalization correspond to X-Y axes in space, as follows: 1) PB space acquired by tactile sensor: Fig.7(a) shows PB space when only tactile sensors were used. Though most objects were not categorized, heavy objects tended to be mapped in upper space. 2) PB space acquired by sound signal: Fig.7(b) shows space when only the sound signal was used. In this space, objects that did not make a sound were mapped at upper right. Although some vague clusters are seen, the sharpness of separation is quite low. 3) PB space acquired by visual data: Fig.7(c) shows space when only visual information was used. In this space, almost all objects were separated, but objects with similar trajectories, such as can or glass and (c) PB space acquired by visual image (d) PB space acquired by all sensor modality Fig. 7. PB spaces acquired by sensor modality. 684 Journal of Robotics and Mechatronics Vol.17 No.6, 2005

5 Figure 8 shows PB space of (a) RNNPB-1 and (b) RNNPB-2. Trained-for objects are shown by white plots and untrained-for objects by black plots. RNNPBparameters 2 corresponding to unknown objects were determined by renewing only parameters without updating synaptic weights (recognition). Renewing the PB value only 1000 times completed recognition. Clusters were self-organized corresponding to all objects in PB space of RNNPB-1. Specifically: 1) Objects that moved easily were mapped at upper left, (a) PB space acquired by RNNPB-1 trained by 8 objects 2) Objects making sounds were mapped in the upper area, 3) Blue objects were mapped at upper right. Fig.8(b) shows that RNNPB-2 acquired almost the same map as Fig.8(a) of RNNPB-1, even though it was trained with data on only four objects. This means that RNNPB- 2 configured PB space with a structure similar to that of RNNPB-1, except for the different area size of each cluster. 5. Discussion (b) PB space acquired by RNNPB-2 trained by 4 objects and recognition results of untrained 4 objects Fig. 8. Comparison of two RNNPBs (generalization analysis). moneybox or pen tray, were not separated. 4) PB space acquired by all sensory modalities: Fig.7(d) shows PB space self-organized when all sensory modalities were used. We confirmed that the RN- NPB acquire clusters for all types of objects Clustering of Unknown Objects We conducted other experiments to confirm our proposal s generalization by recognizing unknown (untrained) objects. In this experiment, we used 8 objects rubber ball, glass, moneybox, pen tray, scrub brush, soft brush, and empty and full shampoo containers. The RNNPB was trained in two different ways. RNNPB-1 was trained using all multimodal sensory data in active sensing. RNNPB-2 was trained using only the rubber ball, glass, scrub brush, and empty shampoo container. For RNNPB-2, moneybox, pen tray, soft brush, and full shampoo container were unknown. Both RN- NPBs were trained 100,000 times Motion Design and Multimodality Motion patterns applicable to different objects are difficult to prepare. Most research on active sensing selected touch or or grasping that focused on tactile and jointdegree sensing. Although these motions guarantee reliable data on shape, size, and weight of objects, such motions require skill for detecting the accurate position of objects to pick them up. Even human infants have difficulty, however, manipulating objects with their hands. We selected the motion of moving or hitting of objects by the robot s arm for active perception. This motion is completed without precise sensor data and many types of dynamic features of objects are extractable, including moving trajectory and sound. Metta et al. used the motion of moving object for active sensing [15], but ignored the important sensor modality of sound. The sound signal reflects many properties of objects such as shape, material, and internal structure. Visual devices alone cannot obtain these properties simultaneously as shown in Fig.7(c). A large number and variety of motion patterns of moving or hitting objects at different speeds and in different directions enabled the extraction of a greater variety of dynamic features of objects Clustering of RNNPB As mentioned in Section 1, most conventional studies on active sensing have dealt with small numbers of objects as the recognition target. Hence it is also difficult to prepare a recognition system that can handle different objects. Typical neural networks for time-sequence data Journal of Robotics and Mechatronics Vol.17 No.6,

6 Ogata, T. et al. processing represented by TDNN [16], for example, require an impractically large number of neurons and learning time for the problem treated in this paper, because it is designed to store all time sequences of sensory data in the input layer. In contrast, the number of neurons in our RNNPB was only 42 because it uses self-organizing contextual information in the context layer Generalization of RNNPB We confirmed that the RNNPB has superior generalization for clustering dynamic sequences. It expresses different objects and their relationships in PB space selforganized through training with a few objects. As mentioned in Section 2, our proposal is more suited for robot learning than stochastic learning because real robot systems have limitations in hardware durability. A deterministic learning system called a mixture of experts represented by MOSAIC [17] also works well in dealing with multiple dynamic patterns (attracters). This usually consists of several dynamic recognizers that categorize and learn target sequences individually (local expression). In contrast, the RNNPB acquires multiattracters overlapping in a single network by changing parameters that represent the boundary condition (distributed expression). In the RNNPB, all neurons and synaptic weights participate in representing all trained patterns. In local expression, interference is minimized between patterns because it allocates a novel pattern in an additional recognizer. In a distributed expression, however, memory interference occurs because memory shares network resources. Due to embedding multiple attractors in a distributed network, we attained a global structure that handles learned patterns and unknown (unlearned) patterns. We believe this is why the RNNPB shows generalization in recognizing unknown objects in Section Conclusions and Projected Work We have proposed active sensing method a humanoid robot with a recurrent neural network to solve the problem of object recognition. Specifically, we trained an RNNPB model with only 42 neurons with data on sounds, trajectories, and tactile senses generated while the robot was moving or hitting an object with its hands. The clusters of 20 types of objects was self-organized in PB space. Experiments using unknown (untrained) objects demonstrated that our proposal configures these unknown objects appropriately in PB space, which proves its generalization. An interesting challenge for projected work is to achieve robot motion planning using our proposal. The targets of our experiments in this paper were different objects but the motion pattern was the same in all active sensing. Our proposal easily shifts its target from different objects to different motions. The RNNPB configuration can be redesigned for treating motor output easily. We expect that our robot will be able to generate arm motion by using RNNPB output. The RNNPB could, for example, associate arm motion patterns with observed object trajectories and sounds. This association was related to the discussion of imitation based on behavioral primitives corresponding to PB in our study. Acknowledgements This research was supported by the RIKEN Brain Science Institute and the Scientific Research on Priority Areas: Informatics Studies for the Foundation of IT Evolution. References: [1] R. Bajcsy, Active Perception, IEEE Proceedings, Special issue on Computer Vision, Vol.76, No.8, pp , [2] K. Noda, M. Suzuki, N. Tsuchiya, Y. Suga, T. Ogata, and S. Sugano, Robust modeling of dynamic environment based on robot embodiment, IEEE ICRA 2003, pp , [3] A. Arsenio, and P. Fitzpatrick, Exploiting cross-modal rhythm for robot perception of objects, Int. Conf. on Computational Intelligence, Robotics, and Autonomous Systems, [4] R. Fukano, Y. Kuniyoshi, T. Kobayashi, and T. Otani, Statistical Manipulation Learning of Unknown Objects by a Multi-Fingered Robot Hand, Humanoids 2004, paper #65, [5] P. Dario, M. Rucci, C. Guadagnini, and C. Laschi, Integrating Visual and Tactile Information in Disassembly Tasks, Int. Conf. on Advanced Robotics, pp , [6] T. Kohonen, Self-Organizing Maps, Springer Series in Information Science, Vol.30, Springer, Berlin, Heidelberg, New York, [7] L. Lin, and T. Mitchell, Efficient Learning and Planning within the Dynamic Framework, SAB 92, pp , [8] J. Tani, and M. Ito, Self-Organization of Behavioural Primitives as Multiple Attractor Dynamics: A Robot Experiment, IEEE Transactions on SMC Part A, Vol.33, No.4, pp , [9] T. Ogata, M. Matsunaga, S. Sugano, and J. Tani, Human Robot Collaboration Using Behavioral Primitives, IEEE/RSJ IROS 2004, pp , [10] M. Jordan, Attractor dynamics and parallelism in a connectionist sequential machine, Eighth Annual Conference of the Cognitive Science Society (Erlbaum, Hillsdale, NJ), pp , [11] D. Rumelhart, G. Hinton, and R. Williams, Learning internal representation by error propagation, in D.E. Rumelhart and J.L. McLelland (editors), Parallel Distributed Processing (Cambridge, MA: MIT Press), [12] H. Ishiguro, T. Ono, M. Imai, T. Maeda, T. Kanda, and R. Nakatsu, Robovie: an interactive humanoid robot, Int. Journal of Industrial Robotics, Vol.28, No.6, pp , [13] T. Miyashita, T. Tajika, K. Shinozawa, H. Ishiguro, K. Kogure, and N. Hagita, Human Position and Posture Detection based on Tactile Information of the Whole Body, IEEE/RSJ IROS 2004 Work Shop, [14] R. Pfeifer, and C. Scheier, Understanding Intelligence, Cambridge, MA: MIT Press, [15] G. Metta, and P. Fitzpatrick, Better Vision through Manipulation, Adaptive Behavior, Vol.11, No.2, pp , [16] A. Waibel, T. Hanazawa, K. Hinton, K. Shikano, and K. Lang, Phoneme Recognition Using Time-Delay Neural Networks, ATR Technical Report TR , [17] M. Haruno, D. Wolpert, and M. Kawato, MOSAIC model for sensorimotor learning and control, Neural Computation 13, pp , Journal of Robotics and Mechatronics Vol.17 No.6, 2005

7 Tetsuya Ogata Associate Professor, Graduate School of Informatics, Kyoto University Jun Tani Laboratory Head, Brain Science Institute, RIKEN Brief Biographical History: Research Associate, Waseda University Research Scientist, Brain Science Institute, RIKEN Graduate School of Informatics, Kyoto University Main Works: ffl Open-end Human-Robot Interaction from the Dynamical Systems Perspective-Mutual Adaptation and Incremental Learning, Advanced Robotics, VSP and Robotics Society of Japan, Vol.19, No.6, pp , July, Membership in Learned Societies: ffl Information Processing Society of Japan (IPSJ) ffl The Robotics Society of Japan (RSJ) ffl The Japan Society of Mechanical Engineers (JSME) ffl Japanese Society for Artificial Intelligence (JSAI) ffl The Society of Biomechanisms (SOBIM) ffl Society of Systems, Control and Information Engineers (ISCIE) ffl The Institute of Electrical and Electronics Engineers (IEEE) Brief Biographical History: Senior Scientist, Sony Computer Science Laboratory Laboratory Head, Brain Science Institute, RIKEN Main Works: ffl The dynamical systems accounts for phenomenology of immanent time: An interpretation by revisiting a robotics synthetic study, Journal of Consciousness Studies, Vol.11, No.9, pp. 5-24, Hayato Ohba Master Course Student, Graduate School of Informatics, Kyoto University Kazunori Komatani Research Associate, Graduate School of Informatics, Kyoto University Membership in Learned Societies: ffl The Robotics Society of Japan (RSJ) Brief Biographical History: Research Associate, Kyoto University Main Works: ffl User Modeling in Spoken Dialogue Systems to Generate Flexible Guidance, User Modeling and User-Adapted Interaction, Vol.15, No.1, pp , Membership in Learned Societies: ffl Information Processing Society of Japan (IPSJ) ffl Japanese Society for Artificial Intelligence (JSAI) ffl The Association for Computational Linguistics (ACL) Journal of Robotics and Mechatronics Vol.17 No.6,

8 Ogata, T. et al. Hiroshi G. Okuno Professor, Graduate School of Informatics, Kyoto University Brief Biographical History: Principal Researcher, Nippon Telephone and Telegram Corp Professor, Science University of Tokyo Professor, Kyoto University Main Works: ffl Advanced Lisp Technology, Taylor & Francis, Aug., ffl Computational Auditory Scene Analysis, Lawrence Erlbaum Associates, ffl Utilizing the Internet, Iwanami Science Library, ffl Intelligent Programming, Ohm Publisher, Membership in Learned Societies: ffl Association for Computing Machinery (ACM) ffl American Association for Artificial Intelligence (AAAI) ffl Acoustical Society of America (ASA) ffl Information Processing Society of Japan (IPSJ) ffl Japanese Society for Software Science and Technology (JSSST) ffl Japanese Society for Artificial Intelligence (JSAI) ffl Japan Cognitive Science Society (JCSS) ffl Institute of Electrical and Electronics Engineers (IEEE) 688 Journal of Robotics and Mechatronics Vol.17 No.6, 2005