Hanover, NH phone: (603) fax: (603) January 9, 1997

Transcription

1 Task-recongurable Robots: Navigators and Manipulators Keith D. Kotay Daniela L. Rus Department of Computer Science Dartmouth College Hanover, NH phone: (603) fax: (603) January 9, 1997 Abstract Task-recongurable robots consist of a set of one or more identical autonomous modules that can adapt their shape and function to tasks. We describe a module that can function as a climbing robot, a manipulator, or a leg in a multi-legged walker. This module can make autonomous transitions between these states. We present the control algorithms that enable our robot to be a versatile navigator and manipulator and report on our experimental results. 1 Introduction A task-recongurable robot is a set of one or more autonomous modules that has the ability to adapt their function and shape according to task, independently. Each module can function as an autonomous robot. These modules also have the ability to come together as a structure and cooperate to exhibit a desired functionality for the structure as a whole. Thus, such a system has the ability to cooperate as a unit or as a distributed system of robots. In this paper we describe our progress towards this class of robots. Specically, we show how our Inchworm robot [KR96] can be an autonomous navigator of three-dimensional structures, a manipulator, and a multi-legged walker. The function of the Inchworm as an autonomous navigator of three dimensional structures, regardless of the presence, absence, or direction of gravitation was reported in [KR96]. Here we describe how the Inchworm robot can (1) attach the front end to an object, lift the object, and push the object on a straight line; (2) attach the back end to an object and pull that object as it propels itself; (3) come together with several Inchworms to form a multilegged walker that can navigate on rough terrain; and (4) transition autonomously between all these tasks, which illustrates the task-recongurability of the Inchworm robot. Applications of this work include inspection tasks, where several Inchworms can attach themselves to a platform and carry tools and power supplies to the base of a complicated bridge or tower. At this point, the Inchworms can detach and distributively inspect the structure by climbing. In our lab we have experimented with using the Inchworm robot as an autonomous climber, autonomous manipulator, and a globally-controlled leg. The Inchworm robot performed successfully each of these tasks and transitions between them many times, with high reliability. Task-recongurable robots can be viewed a minimalist approach to designing versatile and extensible navigators and manipulators. We advocate designing simple modules that can be shown to be robust, such as the Inchworm robot [KR96]. When these modules are autonomous robots, they can independently organize themselves as robust three-dimensional structures whose geometric congurations determine the resulting robot function. Such structures have builtin redundancy and are thus naturally fault-tolerant. One can imagine building robots that can self-organize as any three-dimensional autonomous structure. The robots described in this paper achieve a small step towards this vision, as they can only transform an Inchworm climber into a manipulator or a multi-legged walker. This paper is organized as follows. We continue with a summary of related work. We then describe 1

2 the Inchworm robot and emphasize its new features. We continue by describing the Inchworm as a climber, manipulator, and leg and present experimental data in each of these cases. 2 Related work Our work draws on previous experiences with navigation algorithms [Lat91], designing self-organizing robots, designing biologically-inspired robots, and designing minimalist robot systems [DJR96, DJR94a]. Related work in self-organizing robots includes robots in which modules are recongurable using external intervention [CLBD92]. In [FK90] a cellular robotic system is proposed to coordinate a set of specialized modules. [Yim93] studies multiple modes of locomotion that are achieved by composing a few basic elements in dierent ways. [Mu94] consider a system of modules that can achieve planar motion by walking over each other due to changes in the polarity of magnetic elds. [PCSC] describes metamorphic robots that can aggregate as two-dimensional structures with varying geometry and implement planar locomotion. Task-recongurable robots are dierent in that each module is autonomous, and the resulting structures can be both manipulators and three dimensional navigators. Related work on biologically-inspired robots includes [Bro89, HNT91, FS96, DRJ95]. Brooks [Bro89] proposes insect intelligences and six-legged walkers. Hirose [HNT91, NH94] describes a quadruped robot that uses suction cups for attachment and can climb on straight surfaces (vertically and horizontally) and make transitions between surfaces with joint angles of 90 degrees. This robot is much heavier than ours (99 pounds vs pounds). Our Inchworm robot is dierent than the previous climbing robots [MD92, Neu94, HNT91, GR90, Nis92] in that it is much smaller, lighter, it needs less workspace to operate and it can handle web structures, as well as solid walls. However, our robot requires a ferrous surface while the robots proposed in [Nis92, HNT91] use suction cups and thus do not have this dependency. 3 Background on The Inchworm The Inchworm is a biologically-inspired robot, designed to imitate the movements of the inchworm caterpillar. This functionality was achieved by creating a light, linear structure make of four sections. The sections are linked with three joints providing three degrees of freedom (see Figure 1). These joints allow the Inchworm to extend and ex. The rst and fourth sections are the feet of the Inchworm. These sections contain the attachment mechanisms, equivalent to the legs and protolegs of the inchworm caterpillar, which allow the Inchworm to adhere to the surface it is traversing and provide the anchoring force needed to support the inchworm walking motion. A fourth degree of freedom is provided by a pivot joint which allows the body of the Inchworm to rotate relative to the attachment mechanism of the rear foot. This allows the robot to turn. This pivot joint is currently being integrated in the robot architecture. The weight of the Inchworm is 566 grams. When fully extended, the length of the robot is 330 millimeters and its height is 80 millimeters. When fully contracted, the length of the robot is 175 millimeters and its height is 160 millimeters. The speed of the robot is 0.75 meters/minute. Figure 1: The Inchworm robot. The body of robot consists of four links. The rst and last link of the robot comprise the feet. Each foot has two 1-inch electromagnets attached to it. The robot has one front IR sensor and a pair (IR sensor, contact switch for touch sensing) attached to each electromagnet. The actuators for the joints are servomotors designed for use in radio controlled aircraft. We are currently using the Airtronics Proportional Retract Servo because of its 170 degree rotation capability. The servo weighs 50 grams and produces 5.3 kg-cm of torque. The attachment mechanism of the robot is comprised of electromagnets. There are two 12-volt electromagnets per Inchworm foot, arranged in-line with the body of the Inchworm. These electromagnets provide enough force to securely anchor a foot and support the weight of the robot when the other 2

3 foot is completely extended. This enables our robot to utilize its entire range of motion. We chose electromagnets because they are easy to use and provide a large amount of attachment force in a small package. However, this choice restricts the motion of the robot to steel surfaces and adds weight to the system. The modular design of the robot allows us to replace the electromagnetic modules (mechanical, electronic, and control) with modules that utilize suction cups or grippers. The Inchworm has two types of sensors: touch sensors and infrared proximity sensors. The touch sensors are Omron model EE-SA105 phototransistor optical switches with spring actuator. Two of these sensors are mounted on each foot of the Inchworm, one at the front of the forward electromagnet and one at the back of the rear electromagnet. The touch sensors are used to establish rm contact of the electromagnets with the surface being traversed. The touch sensors also are used to indicate error conditions such as contact with a surface when it is not expected. The infrared proximity sensors are Omron model EE-SF5 reective photomicrosensors. These sensors are used to implement compliant movements along a surface. Five infrared proximity sensors are used on the Inchworm, four mounted adjacent to the touch sensors oriented downward in order to monitor the surface, and one mounted on the front of the Inchworm facing forward. Low-level hardware control of the Inchworm is achieved by an on-board 8-bit PIC16C73 microcontroller. This microcontroller generates the pulse width modulation signals used to control the servomotors and also reads the robot's sensors. Higher-level control is performed by a remote workstation which communicates with the microcontroller via a 38.4k baud serial link. This gives the Inchworm the properties of a remote-brained system, as outlined in [Ina93]. The Inchworm is powered by lead-acid batteries. The batteries are not located on the body of the Inchworm due to weight constraints. This means that the Inchworm must be connected to the power source with a tether. 4 The Inchworm as an autonomous climber The Inchworm robot can propel itself on arbitrarily oriented surfaces in the absence of geometric models. Its motions are compliantly controlled on-line. The control algorithms are simple, compliant, adaptive, and reliable. A control hierarchy is used to partition the control into low-level skills, task-level primitives, and high-level navigation algorithms. Each level in the hierarchy consists of compositions of the functions in the lower levels. In [KR96] we discussed the skills used to perform low-level control of the Inchworm (attach, detach, level, height, raise, extend, lower-to-contact, and position). Refer to [KR96] for a description of each skill. Skills are composed to implement the Inchworm movement primitives step and transition. The step primitive allows the robot to make one propelling step, and the transition primitive allows the robot to move to a dierent surface. There are currently two forms of the transition primitive, concave-transition and convex-transition, used to perform surface transitions depending on the surface angles as described below. We also implemented a high-level algorithm that allows our robot to navigate a three-dimensional structure without models. In the next four sections, we present descriptions of the the movement primitives and the navigation algorithm (step, concave-transition, and the navigation algorithm are discussed in greater detail in [KR96]). 4.1 Stepping The simplest task level primitive is the step. A step consists of two phases, the extension phase and the contraction phase. During the extension phase, the rear foot is attached to the surface. An extension is accomplished by raising the front foot, extending it forward, and then lowering the foot until it contacts the surface. At this point, the front foot is attached to the surface and the rear foot is detached. The contraction phase then occurs by raising the rear foot, contracting the rear foot to bring it closer to the front foot, and then lowering the foot until it contacts the surface. 4.2 Concave Transitions concave-transition is a primitive that implements a move from one surface to another surface when the relative orientation between the surfaces ranges from 45 to 90 degrees relative to the rst surface. (Transitions to surfaces whose relative orientation is between 1 and 45 degrees are implemented as part of a step as explained in [KR96]). The control algorithm for concave-transition consists of raising the front foot, rotating it until it is parallel to the new surface, extending it far enough so that there is space for the rear foot behind the front foot on the new surface, and then bringing the rear foot from the old surface to the new surface. 4.3 Convex Transitions convex-transition is a new primitive used for moves from one surface to another surface when the relative orientation between the surfaces ranges from 3

4 -45 to -90 degrees relative to the rst surface. (Transitions to surfaces whose relative orientation is between -1 and -45 degrees are implemented as part of a step in the same way as for a concave transition as explained in [KR96]). The control algorithm for convex-transition consists of moving the front foot to the new surface, bringing the back foot close to the corner, extending the front foot further on the surface to make room for the back foot and then bringing the back foot to the new surface (see Figure 2). In order to implement convex-transition, it was necessary to increase the length of the middle two Inchworm segments from 80 millimeters as in [KR96] to the current length of 120 millimeters. This was required because convex-transition needs a greater reach than concave-transition. A consequence of increasing the segment lengths is an increased torque demand on the joint actuators, since at maximum extension the moment arm length is increased. As a result of the segment length increase, our current implementation is restricted to -45 degree transitions. We are investigating more powerful servomotors in order to remove this restriction. 4.4 On-line walking in three dimensions We have implemented an on-line navigation algorithm which implements a greedy strategy where the robot attempts to preserve the initial direction of motion. If the IR sensor pointed forward registers an object (e.g. it returns a small value) it follows that an obstacle or a surface with a relative concave orientation has been encountered. The robot uses concave-transition to attach itself to this new surface and continues travelling. If the rst IR sensor of the bottom of the front foot registers free space (e.g. it returns a large value while the second IR sensor on that foot returns a small value) a surface with a relative convex orientation has been encountered. The robot uses convex-transition to attach itself to this new surface and continues travelling. Otherwise, the robot continues stepping in the current direction. 5 The Inchworm as a manipulator We have recently begun to investigate the capabilities of the Inchworm as an autonomous manipulator. The ability of the robot to attach itself securely to surfaces empowers the Inchworm with the ability to grasp objects with one foot while using the other foot to maintain a stable attachment to a surface. In this way, the Inchworm can exert forces on the grasped object causing it to move. In the following sections we present two examples of Inchworm manipulation: lifting and placing an object, and pulling an object. 5.1 Lifting and placing an object When the Inchworm is performing a concave transition, it must detect the surface in its path and then attach its front foot to that surface. These are the same skills the Inchworm needs to detect and grasp a movable object. We have implemented a lift-and-place algorithm which grasps an object, lifts the object, moves the object forward, lowers the object to the surface, and returns the Inchworm to its original pose. We have used lift-and-place to develop a higherlevel algorithm, sweep, which can move an object in a straight line. sweep repeatedly performs a step until an object is encountered, then positions the robot, invokes lift-and-place, and continues stepping (see Figure 3). When an object is detected by the forward IR sensor during a step movement, the step is terminated and the Inchworm is positioned to perform a grasp. A raise skill with forward compliance is then used to lift the front foot while maintaining a constant distance to the object. Then a position skill is performed to align the bottom of the front foot with the object, followed by an extend skill which terminates on the detection of free space by the forward, downward-pointing IR sensor on the front foot, signifying that the front foot is near the top of the object. A lower-to-contact skill places the electromagnets on the surface of the object and an attach completes the grasp. Once the object is grasped, the raise skill is used to lift and lower the object, and the extend skill is used to move the object forward. A detach releases the object. 5.2 Pulling an object In our previous work, we developed control algorithms which allowed the Inchworm to move autonomously in an unstructured environment. We now present three manipulation primitives which allow the Inchworm to attach itself to an object, move itself and the object, and detach itself from the object. These primitives enable the Inchworm to manipulate objects in its environment. Our current experimental platform is a wheeled object with an attached steel surface at an angle of approximately 45 degrees (see Figure 4). This object rests on a level steel surface, or \oor". The robot manipulates the object by attaching one foot to the steel surface of the object, attaching the other foot to the oor, and then generating a propelling motion which moves the entire system. The manipulation primitives are connect, stride, and disconnect. connect allows the Inchworm to attach its rear foot to the object to be manipulated. This is done using a series of position movements which 4

5 Figure 2: Illustration of the movements performed during a convex transition. Figure 3: Six snapshots taken from a sweep sequence place the rear foot in close proximity to the inclined steel surface, followed by a lower-to-contact movement which moves the rear foot into contact with the surface. attach is then used to activate the attachment mechanism. When the rear foot is attached, the front foot can be raised o the \oor" surface due to the stability of the wheeled platform. This allows the front foot to be placed in the starting point for the stride primitive. The stride primitive is used to move the Inchworm and wheeled platform system across the steel oor. The algorithm begins with an extension of the front foot identical to that in the extension phase of the step algorithm described in Section 4.1. After the front foot is in contact with the surface, the front foot is attached to the steel oor and three position movements are performed which, in turn, push down on the oor raising the small castor ball on the wheeled platform, pull the wheeled platform forward toward the front foot, and lower the castor ball to the oor. This results in a net forward motion of the object. disconnect is a primitive that is used to detach the Inchworm from the manipulation object. It consists of a detach skill which deactivates the attachment mechanism on the rear foot, followed by a raise skill which lifts the rear foot o the steel attachment surface. Two position movements and a lower-to-contact movement return the rear foot to contact with the oor. Using these primitives the Inchworm can perform manipulation tasks on objects such as the wheeled platform (see Figure 4). Due to the lack of turning capability on the Inchworm, manipulation is limited to straight-line motion. We have only implemented a pulling motion, but a pushing motion is also feasible. 5.3 Transitions between navigation and manipulation The design of our control code enables autonomous transitions between manipulation and navigation tasks. This is demonstrated by the use of a navigation primitive, step, and a manipulation primitive, lift-and-place, within the sweep algorithm. sweep autonomously transitions between navigation and manipulation primitives in the process of moving an object. In the object pulling task, the Inchworm uses manipulation primitives to move the object but once the robot has detached itself from the object it can switch 5

6 Figure 4: Eight snapshots taken from a pull sequence followed by a step. to navigation primitives to move away from the object as shown in Figure 4. The ability to autonomously transition between navigation and manipulation tasks gives the Inchworm the exibility to navigate or manipulate as circumstances dictate. 6 Multi-legged walker In Section 5.2 we showed how the Inchworm can manipulate an object by pulling it. Multiple Inchworms can run the same algorithm in synchrony. This would be useful when the weight of the object being pulled is too large to be moved by one Inchworm alone. If the frictional forces between the object being pulled and the surface are high, multiple Inchworms can attach themselves to the object and lift it to suspend the object in air. The Inchworms can then propel themselves in synchrony to relocate the object. This resulting structure can be thought of as a multi-legged walker, where each inchworm functions as one leg. At least 4 inchworms are needed to implement such a structure (see Figure 5). Three Inchworms are needed to be in contact with the surface at all times to stably support the object. A four-legged structure can walk by sequentially allowing each Inchworm to perform the stride described in Section 5.2. A six-legged walker can walk by using the tripod gate. Other interesting gates can be formulated when more than six legs are involved. As described in Section 5.3, the Inchworm can autonomously transition between navigation and manipulation tasks. This capability enables individual Inchworms to independently navigate to the walker structure and then use manipulation primitives to attach to and move the structure. Similarly, Inchworms can detach from the walker structure and switch to navigation primitives in order to perform individual tasks. Figure 5: A four-legged walker composed of four Inchworm robots. 6.1 Globally-controlled leg Although we have not yet implemented this algorithm because we have not completed the construction of all four Inchworms, we have implemented the function of one leg. The stride movement performed as part of the pulling task executes a motion which propels an attached object forward. This one-leg control can be replicated on three other Inchworms. Here we assume that a global controller coordinates the specic roles of the Inchworms in this structure (that is, tells the inchworms when to stay xed to support the object and when to execute the stride). Of course, many interesting challenges arise in a distributed model, where Inchworms have local control and some communication mechanism. We plan to address the control of a distributed set of Inchworms in our future work. 6

7 7 Experiments We implemented the sweep algorithm to test the lift-and-place primitive, and the pull algorithm to test the connect, stride, and disconnect primitives. All experiments were conducted on a level steel surface. The object used in the sweep experiments was a balsa wood block with a steel plate attached to one side. The weight of the block is 116 grams. A successful lift-and-place means that the object was grasped, lifted, moved forward, and lowered. Each successful lift-and-place also implies that the object was successfully detected and the Inchworm was correctly positioned for the lift-and-place operation, however object detection and Inchworm positioning are part of the sweep algorithm not lift-and-place. In this experiment, all objects detected in front of the Inchworm were assumed to be movable objects. As shown in Figure 6, in 115 trials lift-and-place was successful over 92 percent of the time. In the straight-line pulling experiment, the Inchworm uses connect to attach itself to a wheeled platform, pulls the platform using two stride operations, detaches from the object using disconnect, and then performs two step operations. The results of 35 runs of this experiment are shown in Figure 6. As the data shows, the primitives are very reliable. It is worth noting, however, that the connect primitive is sensitive to the initial positions of the robot and the object. For the above experiment, the initial positions of the robot and the object were set by hand. The reason for this is that the Inchworm does not have a backward facing sensor to detect the position of the object. We intend to add a sensor to overcome this limitation. Figure 6 also includes data for step and the globally-controlled leg algorithm. These tasks were components of the straight-line pulling experiment. Each pulling experiment included two step tasks, and this data was added to the data presented in [KR96]. The result is 200 successful steps out of 210 tries. The globally-controlled leg data consists of data for the stride component of the pulling experiment. This was done because a stride consists of the leg motion which would be used for each Inchworm in the multilegged walker. 8 Discussion and Future work We have described a task-recongurable robot that can accomplish complex multi-modal navigation and manipulation tasks in three dimensional environments. The experimental data demonstrates that this is a feasible and robust way of creating versatile Task Tries Success Reliability Lifting and Placing % Straight-line Pulling % Step % Globally-Controlled Leg % Figure 6: This table contains reliability data for the lifting and placing, and straight-line pulling experiments. Data for the globally-controlled leg entry is from the straight-line pulling experiment. Data for the step entry is from the straight-line pulling experiment and [KR96]. robots. At the moment we have experimented with one module only, but we are currently building ve more Inchworm robots. Our research agenda is to develop autonomous robots that are simple, exible, and extensible. Our long term goal is to use this simple platform to study cooperation among Inchworm robots and also between Inchworm robots and other robots in our lab. We want to develop an extensible fault-tolerant architecture that would allow robots to autonomously come together and move apart in order to cooperate in distributed tasks. We plan to build a total of 6 Inchworm robots and develop a mechanism that allows them to connect into one single unit that could travel by using each individual Inchworm as a leg. This six-legged walker will demonstrate a redundant locomotion property in that a leg can be discarded if it malfunctions. In addition, legs will be able to recongure themselves as pairs of manipulators, allowing the robot system to alter its environment. Acknowledgements This paper describes research done in the Dartmouth Robotics Laboratory. Support for this work was provided through the NSF CAREER award IRI Support for our research was also provided by Microchip Inc., the Motorola University Support Program, Omron Inc., and RWI Inc. We are grateful for it. We would also like to thank David Collins and Dale Lihl for their assistance in our research eort. References [Bro89] R. Brooks, A robot that walks: emergent behaviors from a carefully evolved network, in Proceedings of the IEEE Conference on Robotics and Automation, Scottsdale, [CLBD92] R. Cohen, M. Lipton, M. Dai, and B. Benhabib, Conceptual design of a modular robot, Journal of Mechanical Design, March 1992, pp [DRJ95] R. Desai, C. Rosenberg, and J. Jones, Kaa: an autonomous serpentine robot utilizes behavior con- 7

8 trol, In Proceedings of the 1995 International Conference on Intelligent Robots and Systems, Pittsburgh, [DJR96] B. Donald, J, Jennings, and D. Rus, Minimalism + Distribution = Supermodularity, in Journal of Experimental and Theoretical Articial Intelligence, (to appear) [DJR94a] B. Donald, J. Jennings, and D.Rus. Information invariants for distributed manipulation, The First Workshop on the Algorithmic Foundations of Robotics, eds. K. Goldberg, D. Halperin, J.-C. Latombe, and R. Wilson, pages 431{459, [FS96] A. Fernworn and D. Stacey, Inchworm Mobility - Stable, Reliable and Inexpensive, In Proceedings of the Third IASTED Internation Conference on Robotics and Manufacturing, Cancun, [FK90] T. Fukuda and Y. Kawauchi, Cellular robotic system (CEBOT) as one of the realization of self-organizing intelligent universal manipulator, in Proceedings of the 1990 IEEE Conference on Robotics and Automation, pp [GR90] V. Gradetsky and M. Rachkov, Wall climbing robot and its applications for building construction, Mechatronic Systems Engineering 1: , Kluwer Academic Press, [HNT91] S. Hirose, A. Nagakubo, and R. Toyama, Machine that can walk and climb on oors, walls, and ceilings, in Proceedings of the International Conference on Advances in Robotics, Pisa, , [Ina93] M. Inaba, Remote-Brained Robotics: Interfacing AI with real world behaviors, in Robotics Research: The Sixth International Symposium, Hidden Valley, [KR96] K. Kotay and D. Rus, Navigating 3d steel web structures with an inchworm robot, in Proceedings of the 1996 International Conference on Intelligent Robots and Systems, Osaka, [Lat91] J. C. Latombe, Robot Motion Planning, Kluwer Academic Publishers [MD92] A. Madhani and S. Dubowsky, Motion planning of multi-limb robotic systems subject to force and friction constraints, in Proceedings of the IEEE Conference on Robotics and Automation, Nice, [Mu94] S. Murata, H. Kurokawa, and Shigeru Kokaji, Self-assembling machine, in Proceedings of the 1994 IEEE International Conference on Robotics and Automation, San Diego, [Neu94] W. Neubauer, A spider-like robot that climbs vertically in ducts or pipes, in Proceedings of the 1994 International Conference on Intelligent Robots and Systems, Munich, [NH94] A. Nagakubo and S. Hirose, Walking and running of the quadruped wall-climbing robot, in Proceedings of the IEEE Conference on Robotics and Automation, pp , San Diego, [Nis92] A. Nishi, A biped walking robot capable of moving on a vertical wall, Mechatronics, vol2, no. 6, pp , [PCSC] A. Pamecha, C-J. Chiang, D. Stein, and G. Chirikjian, Design and implementation of metamorphic robots, in Proceedings of the 1996 ASME Design Engineering Technical Conference and Computers in Engineering Conference, Irvine, CA [Yim93] M. Yim, A recongurable modular robot with multiple modes of locomotion, in Proceedings of the 1993 JSME Conference on Advanced Mechatronics, Tokyo, Japan