Uri Kartoun*, Helman Stern*, Yael Edan*, Craig Feied**, Jonathan Handler**, Mark Smith**, Michael Gillam**

*Department of Industrial Engineering and Management, Ben-Gurion University of the Negev

Be’er-Sheeva, 84105, Israel

{kartoun, helman, yael}@bgu.ac.il

**Institute for Medical Informatics, Washington Hospital Center

110 Irving St., Washington DC, NW, 20010, U.S.A.

{cfeied}@ncemi.org

Download Paper (pdf)

ABSTRACT

This paper presents a method for autonomous recharging of a mobile robot, a necessity for achieving long-term robotic activity without human intervention. A recharging station is designed consisting of a stationary docking station and a docking mechanism mounted to an ER-1 Evolution Robotics robot. The docking station and docking mechanism serve as a dual-power source, providing a mechanical and electrical connection between the recharging system of the robot and a laptop placed on it. Docking strategy algorithms use vision based navigation. The result is a significantly low-cost, high-entrance angle tolerant system. Iterative improvements to the system, to resist environmental perturbations and implement obstacle avoidance, ultimately resulted in a docking success rate of 100 percent over 50 trials.

KEYWORDS: Robotic mechanisms, robot control, autonomous navigation, machine vision

1. INTRODUCTION

     Mobile robots are being designed to interact increasingly with human environments, working with and around humans on a daily basis. To be considered of any use, these robots must exhibit some form of self-sustainability. In addition to being robust in their physical design as well as control methodology, such robots must be capable of long-term autonomy. Energy is of great concern, and without it the robot will become immobilized and useless [1].

     A custom recharging station designed for ActivMedia Pioneer robots is described in [2]. The station emits an infrared signal that is detected by a ring of seven infrared receivers located at the back of the robot. The robot docks into the recharging station by backing up so that the charging pins make contact with the recharging station. Aside from the Pioneer-based recharging system, there are other efforts under way to develop recharging capabilities to achieve long-term autonomous mobile robot control. Examples include [3] and [4], which describe the same recharging system with increased functionality. 94.26 percent docking success is addressed in [4], which describes the results of repetitive docking over the course of a week using a robot similar to the Pioneer. Commercially available robots examples include: (i) the Helpmate robot [5] having been installed in hospitals worldwide, and (ii) a cleaning machine equipped with a Siemens Corporation navigation system [6]. Chips, an autonomous robotic technology XR4000-based was implemented on three robots in museums providing tours and covering travel distance of hundreds of kilometers [7]. In order to achieve true self-reliance, any of the three robots must be able to recharge themselves when necessary. This was accomplished by using a simple 3D fiducial, aligned with an electrical outlet that provided both translational and rotational position feedback. Using this marker, the robots had demonstrated reliable positioning to an accuracy of 1.5 mm using visual position servoing. The entire docking process, including moving over a distance of four meters to the outlet and then fine-servoing for the insertion process, took less than three minutes.

     97 percent success rate for docking within average time of 30.5 seconds using a Pioneer robot was achieved by [1, 8]. The docking station is a freestanding passive device with two degrees of freedom (DOF), providing the necessary compliance for multiple docking conditions. The robot docking mechanism is a passive one DOF assembly attached to the back of the robot. A contact switch is connected to the underside of the docking mechanism, controlling the power connection to the robot’s batteries. An IR-LED detector is mounted to the top deck of the robot directed toward its back.

     Our focus here is on the development of long-term autonomous capabilities that enable an ER-1 robot to recharge itself without assistance. We developed a recharging station that allows the robot to recharge autonomously. The recharging station’s physical design is described in two parts; (i) the docking station, and (ii) the robot-laptop docking mechanism. Our recharging and docking strategy is presented, including an explanation of the tasks developed to control the docking navigation procedures. The system can be easily adapted to different indoor environments, as well as multiple robots.

    The physical design is presented in section 2. Recharging and docking strategy describing the navigation control and vision algorithms is presented in section 3. In section 4 experiments and results are described. The paper ends in section 5 with conclusions and some directions for future work.

2. PHYSICAL DESIGN

     Our system is based around an ER-1 Evolution Robotics mobile robot. This robot is a three-wheeled platform equipped with an IBM ThinkPad laptop and a Logitech QuickCam Pro 4000 color camera supporting autonomous control capabilities. The ER-1 rechargeable power module equipped with a 12V, 5.4A battery can keep it running for 2.5 hours. In addition power should be supplied to the laptop battery.

     The design presented consists of two parts: a docking station (Figure 1(a)) and a robot-laptop docking mechanism (Figure 1(b)). Design factors include a large tolerance window for docking to increase the probability of docking success over a wide range of robot entry angles. The robot may enter the docking station with a high probability of success within a total entry angle of 45°. Minimal modification to the ER-1 robot is desired, reducing any interference with current robot capabilities.

(a) Docking station	(b) Robot-laptop docking mechanism

Figure 1. System physical design

2.1 Docking Station

     The docking station (Figure 1(a)) is a stationary fixture that provides a connection point for the robot-laptop’s docking mechanism described in section 2.2. The ER-1 charger as well as the laptop power supplier are connected to the docking station providing the necessary power. The docking station is LEGO-based and built of cubes assembled into two surfaces. Four aluminum surfaces interlace into the structure. The upper two panels are connected to a 16VDC adapter for recharging the robot laptop. The two lower panels are connected to a 12VDC power charger for recharging the ER-1 battery.

2.2 Robot-Laptop Docking Mechanism

     The robot-laptop docking mechanism (Figure 1(b)) is mounted to the ER-1 robot. The mechanism consists of two identical four flexible spring structures located on the left and on the right of the back of the robot. The upper two springs of each structure are connected to the laptop and the lower two are connected to the robot battery. Encounter of any one of the two structures with the docking station results in a successful docking and beginning of recharging.

3. RECHARGING AND DOCKING STRATEGY

3.1 Navigation Control

     The system was tested using two robotic configurations, each one was tested for a navigation task in a different environment: (i) "Docking Room I" - the robot-laptop docking mechanism was mounted opposite to the robot motion while the camera was facing toward (Figure 2(a)), and (ii) "Docking Room II" - both the robot-laptop docking mechanism and the camera were mounted toward the robot motion (Figure 2(b)).

(a) Robot-laptop docking mechanism mounted opposite to the robot motion	(b) Robot-laptop docking mechanism mounted toward the robot motion

Figure 2. Overall views over the docking system and the ER-1 robot

     The control is built using several tasks: (i) the Drive-Toward-Object task makes the ER-1 to track an image while moving and correcting its location to drive toward this image; (ii) the Turn-To-Object task operates in three modes: (a) search; (b) hover, and (c) track. In this mode the robot tries to turn toward the target to face it directly; (iii) the Obstacle-Avoidance inhibits the Drive-Toward-Object task in order to directly control the motors and stop the robot when the range between the robot and an object (e.g., human) is sufficiently small, and (iv) the Proceed-And-Scan task triggers the robot to move short distance forward (e.g., 5 cm) and scan its surroundings (e.g., rotate 5° to the left and to the right) repeatedly. This task repeats itself when the robot can not identify any landmarks of an image till the target is found and assures that in cases of low target visibility or camera trembling due to nonuniform driving the robot arrives to its destination. Flowchart describing the navigation strategy is shown in Figure 3.

Figure 3. Navigation flowchart

3.2 Object Recognition

     The object recognition algorithm is based on extracting salient features from an image or an object [9, 10]. Local descriptors [e.g., 11] are commonly employed in a number of real-world applications [e.g., 12] because they can be computed in real-time, resistant to partial occlusions, and are relatively insensitive to changes in viewpoint.

     It is stated in [13] that "Hessian-Laplace regions [11, 14] are invariant to rotation and scale changes. Points are localized in space at the local maxima of the Hessian determinant and in scale at the local maxima of the Laplacian-of-Gaussians. This detector is similar to the approach described in [10] which localizes points at local scale-space maxima of the difference-of-Gaussian (DoG). Both approaches detect the same blob-like structures. However, Hessian-Laplace obtains higher localization accuracy in scalespace, as DoG also responds to edges and detection is unstable in this case. The scale selection accuracy is also higher than in the case of the Harris-Laplace detector".

     The main strength of the object recognition algorithm used here [9, 10] lies in its ability to provide reliable recognition in realistic environments where lighting conditions change dramatically. It is stated in [9] that "the time it takes to process and match each training or recognition image is about 0.8 seconds on a Pentium III 600MHz computer. About half of the time is devoted to image acquisition and feature detection, while the remaining time is used for all aspects of matching and model formation". The algorithm consists of two stages: (i) training - accomplished by capturing an image of an object or a scene, and (ii) recognition of local features for each image the robot encounters. A small subset of those features and their interrelation identifies a pose and a distance of the image.

     Each feature is uniquely described by the texture of a small window of pixels around it [10]. The model of an image consists of the coordinates of all these features along with each feature’s texture description. When the algorithm attempts to recognize objects in a new image, it first finds features in the new image. It then tries to associate features in the new image with all the features in the database of models. This matching is based on the similarity of the feature texture. If eight or more features in the new image have good matches to the same database model, that potential image match is refined. The refinement process involves the computation of an affine transform between the new image and the database model, so that the relative position of the features is preserved through the transformation. The algorithm outputs all object matches for which the optimized affine transform results in a small root-mean-square pixel error between the features found in the new image, and the corresponding affine-transformed features of the original model.

     As stated in [9], the similarity transform between two images gives the mapping of a model point  to an image point  in terms of an image scaling,  , an image rotation, , and an image translation, :

which can be written as a linear form collecting the unknown similarity transform parameters into a vector:

This linear system can be written as

The least-squares solution for the parameters  are determined by solving the corresponding normal equations:

which minimizes the sum of squares of the distances from the projected model locations to the corresponding image locations.

By using the solution of  the average error,  between each projected model feature and image feature can be estimated:

where  is the number of rows in matrix . As each new training image arrives, it is matched to the previous model views. The result is that training images that are closely matched by a similarity transform are clustered into model views (number of views is determined dynamically) that combine their features for increased robustness. Otherwise, the training images form new views in which features are linked to their neighbors. The result is that additional training images continue to contribute to the robustness of the system by modeling feature variation without leading to a continuous increase in the number of view models [9].

4. EXPERIMENTS AND RESULTS

     Test conditions were developed, consisting of a closed environment for the robot to maneuver and two recharging tasks. When the ER-1 is placed at anyone of "Docking Room I" or "Docking Room II" entrances (Figure 5) it automatically starts its navigation. Performance is based on the success or failure of the docking and recharging operation.

     The object recognition algorithms were integrated with the navigation control strategies. The image databases were trained using three different images for each one of the robot-laptop mechanism configurations applied in the docking rooms (Figure 2). An example for features extracted from the docking station image is shown in Figure 4. Figure 4(a) shows the corresponding 724 features (white spheres) overlaid extracted from the docking station image. Figure 4(b) shows the corresponding features overlaid (white spheres) as well as the set of matched features (dark spheres) extracted from an image captured by the camera when the robot moves toward the docking station (28 matched features and 540 features in total). It is noted the height of the robot is taken into account when the images are collected; the height of the camera placed on the robot was set to 65 cm above the floor and was constant during the experiments, both during training and navigation. If it is desired to place the camera in a different height, the training stage should be repeated.

(a) Features extracted from the original docking station image	(b) Features extracted when the ER-1 moves toward the docking station image

Figure 4. Example for features extracted from the docking station image

To determine the performance capabilities of our system 50 trials were conducted for each one of the docking rooms. During "Docking Room I" experiments, in 25 of the trials the robot was placed at the west entrance and in the other 25 trials the robot was placed at the north entrance. Figure 5(a) describes "Docking Room I" structure and shows odometry results of six sample trials (at three of the trials noted as spheres, the robot starts the docking procedure from the west entrance and at the other three trials noted as diamonds it starts from north). Results of the 50 trials showed a 98 percent success rate for mechanical docking, and a 96 percent success rate for the electrical recharging operation. We distinguish between mechanical docking and electrical recharging separately due to the differences involved to complete each operation. Mechanical docking is based on whether or not the robot entered or missed the docking station. The electrical docking indicates whether a recharging procedure has been initiated. Both mechanical and electrical dockings are necessary to allow power to flow to the batteries from the chargers. The one mechanical docking failure in "Docking Room I" during the 50 trials was due the reason that a person stood in the robot’s path during the docking procedure and the robot activity was distracted. The two electrical recharging failures were due to unseen dirt and dust that covered the docking station’s aluminum surfaces and caused nonconductivity between the contacts.

(a) "Docking Room I"	(b) "Docking Room II"

Figure 5. Top view over the robotic environments and odometry results

     In order to improve our dock success rate achieved at the "Docking Room I" experiment, the Proceed-And-Scan task was integrated with "Docking Room II" navigation algorithm. Furthermore, another capability of recognizing whether docking is successful was added to the system; the robot keeps measuring the battery percentage left. An increase of at least three percent in the robot battery level indicates that charging has been established. Figure 5(b) describes "Docking Room II" structure and shows odometry results of three sample trials noted as spheres. Results of 50 trials showed a 100 percent success rate for both mechanical and electrical dockings.

     Our tests showed that the recharging and docking control algorithms were capable of positioning the robot with an accuracy of approximately 15° relative to the docking station in both docking room experiments. This falls within the allowable tolerance zone of the docking station. The time associated with the entire docking procedure was measured for each trial. This is the time from which the robot enters the room to the point it successfully docks mechanically and electrical contact is initiated. We determined that in "Docking Room I", the average docking time spanned approximately 57.4 seconds for the west entrance with a standard deviation of 2.4 within an average distance of 400.3 cm and 51.6 seconds for the north entrance with a standard deviation of 3.5 within an average distance of 410 cm. A best case scenario resulted in 55.3 seconds for the west entrance and 46.2 seconds for the north entrance dock time. Our worst case trial in "Docking Room I" experiment where the ER-1 started from the west entrance resulted in a 66.4 seconds dock time due to the reason that a human passed through the robot trajectory and caused it to wait till she left the room. In "Docking Room II", the average docking time spanned approximately 85.23 seconds with a standard deviation of 20.84 within an average distance of 768.2 cm. A best case scenario resulted in 60 seconds. Our worst case trial in this experiment resulted in a 174.52 seconds dock time due to the reason that several curious people gathered around the robot surroundings and blocked its way occasionally.

5. CONCLUSIONS AND FUTURE WORK

     We successfully developed a recharging system and navigation control algorithms that have been integrated with an ER-1 robot, thereby providing the capabilities necessary for long-term autonomous experimentation. The docking station concept presented is not limited to any specific robot. Our recharging system was tested resulting in a 100 percent mechanical and electrical success rate for docking, within less than 1.5 min average dock time for each dock. Although the robot met obstacles during its navigation, the algorithm coped with all of them and worked continuously. With comparison with the reviewed docking stations mentioned our system allows considerably high degree of entrance tolerance (45°). Although the docking strategy described in [1] resulted in better average docking times, our simply designed system provided a remarkable percent success of mechanical and electrical docking rate within short dock times while tested in environments with external interruptions (e.g., people). In addition, no double recharging system (i.e., recharging a computer in addition to a robot battery) is discussed in literature. We note that our system achieves high dock rate success in comparison with other systems due to: (i) the unique hardware we installed on the back of the robot, which uses two flexible spring structures as part of its robot-laptop docking mechanism instead of only one, a fact that reduces the probability of mechanical docking failures, and (ii) a design that allows relatively large positioning errors due to its relatively large size. Additional trials are planned to prove the robustness and repeatability of the system over extended periods of time, under various conditions and with different robots.

6. ACKNOWLEDGEMENTS

     This work is partially supported by the Paul Ivanier Center for Robotics Research and Production Management, Ben-Gurion University of the Negev and by the Rabbi W. Gunther Plaut Chair in Manufacturing Engineering. The assistance of Mr. Robert Irving from the Institute for Medical Informatics is greatly appreciated.

7. REFERNCES

[1] M.C. Silverman, D. Nies, B. Jung, and G.S. Sukhatme, "Staying alive: a docking station for autonomous robot recharging," IEEE International Conference on Robotics and Automation, vol. 1, pp. 1050-1055, Washington DC, 2002.

[2] F. Michaud, J. Audet, D. Letourneau, L. Lussier, C. Theberge-Turmel, and S. Caron, "Autonomous robot that uses symbol recognition and artificial emotion to attend the AAAI conference, " AAAI Robot Competition Technical Report, 2000.

[3] H. Yasushi and Y. Shin’ichi, "Robust navigation and battery re-charging system for long-term activity of autonomous mobile robot," Proceedings of the 9^th International Conference on Advanced Robotics, pp. 297-302, 1999.

[4] H. Yasushi and Y. Shin’ichi, "A first stage experiment of long term activity of autonomous mobile robot - result of repetitive base-docking over a week," Proceedings of The 7^th International Symposium on Experimental Robotics (ISER2000), pp. 235-244, 2000.

[5] S. King and C. Weiman, "Helpmate autonomous mobile navigation system," Proceedings of SPIE Conference on Mobile Robots, vol. 2352, pp. 190-198, Boston, MA, 1990.

[6] H. Endres, W. Feiten and G. Lawitzky, "Field test of a navigation system: autonomous cleaning in supermarkets," Proceedings of the 1998 IEEE International conference onRobotics and Automation, pp. 1779-1781, Leuven, Belgium, 1998.

[7] I.R. Nourbakhsh, C. Kunz and T. Willeke, "The mobot museum robot installations: a five year experiment," Proceedings of the 2003 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2003), vol. 3, pp. 3636 - 3641, 2003.

[8] M.C. Silverman, D. Nies, B. Jung, and G.S. Sukhatme, "Staying alive longer: autonomous robot recharging put to the test," Center for Robotics and Embedded Systems, CRES-03-015 Technical Report, University of Southern California, 2003.

[9] D. Lowe, "Local feature view clustering for 3D object recognition," Proceedings of the 2001 IEEE/RSJ, International Conference on Computer Vision and Pattern Recognition, Kauai, Hawaii, U.S.A., pp. 682-688, 2001.

[10] D. Lowe, "Object recognition from local scale-invariant features," Proceedings of the International Conference on Computer Vision, pp. 1150-1157, 1999.

[11] D. Lowe, "Distinctive image features from scale-invariant keypoints," International Journal of Computer Vision, vol. 2, no. 60, pp. 91-110, 2004.

[12] M. E. Munich, P. Pirjanian, E. Di Bernardo, L.Goncalves, N. Karlsson, and D. Lowe, "Break-through visual pattern recognition for robotics and automation," Proceedings of the 2005 IEEE International Conference on Robotics and Automation, Barcelona, Spain, 2005.

[13] K. Mikolajczyk and C. Schmid, "A performance evaluation of local descriptors," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 27, no. 10, pp. 1615-1630, October, 2005.

[14] K. Mikolajczyk and C. Schmid, "Scale and affine invariant interest point detectors," International Journal of Computer Vision, vol. 1, no. 60, pp. 63.86, 2004.