機(jī)器人在自然地形下爬行【中文6940字】【PDF+中文WORD】,中文6940字,PDF+中文WORD,機(jī)器人,自然,地形,爬行,中文,6940,PDF,WORD Climbing Robots in Natural Terrain Timothy Bretl, Teresa Miller, and Stephen Rock Jean-Claude Latombe Aerospace Robotics Lab Robotics Laboratory Department of Aeronautics and Astronautics Computer Science Department Stanford University, Stanford, CA 94305 Stanford University, Stanford, CA 94305 {tbretl, tgmiller, rock}@sun-valley.stanford.edu latombe@cs.stanford.edu Keywords Motion planning, climbing, robotics, legged robots, high-risk access, natural terrain. Abstract This paper presents a general framework for plan-ning the quasi-static motion of climbing robots. The framework is instantiated to compute climbing motions of a three-limbed robot in vertical natural terrain. An example resulting path through a large simulated environment is presented. The planning problem is one of five fundamental challenges to the development of real robotic systems able to climb real natural terrain. Each of the four other areas—hardware design, control, sensing, and grasping—is also discussed. 1 Introduction The work described in this paper is part of an effort to develop critical technologies that will enable the design and implementation of an autonomous robot able to climb vertical natural terrain. To our knowl-edge, this capability has not been demonstrated previously for robotic systems. Prior approaches have dealt with artificial terrain, either using special “grasps” (e.g., pegs, magnets) adapted to the terrain’s surface or exploiting specific properties or features of the terrain (e.g., ducts and pipes) [1-12]. Developing this capability will further our under-standing of how humans perform such complex tasks as climbing and scrambling in rugged terrain. This may prove useful in the future development of sophisticated robotic systems that will either aid or replace humans in the performance of aggressive tasks in difficult terrain. Examples include robotic systems for such military and civilian uses as search-and-rescue, reconnaissance, and planetary exploration. Many issues need to be addressed before real robots can climb real, vertical, natural terrain. This paper considers five of the most fundamental of these issues: hardware design, control, sensing, planning, and grasping. One of these issues in particular, the motion-planning problem, is described in more detail. A general framework for climbing robots is presented and this framework is instantiated to compute climbing  Fig 1. A three-limbed climbing robot moving vertically on natural surfaces. motions of the three-limbed robot shown in Figure 1. Simulation results are shown for the robot in an example vertical environment. 2 Motivation The results of research in this area will benefit a number of applications and have implications for several related research areas. 2.1 Applications This paper is motivated by a need for robotic sys-tems capable of providing remote access to high-risk natural environments. There are many terrestrial applications for these systems, such as search-and-rescue, cave exploration, human assistance for rock and mountain climbing, and tactical urban missions. Each of these applications requires climbing, descending, or traversing steep slopes and broken terrain, and thus involves consider-able human risk. Several space applications could also benefit from these aggressive robotic systems. For example, sites on Mars with potentially high science value have been identified on cliff faces [13]. Often, it is neither practical nor feasible for flying robots to access these locations. Therefore, to reach these sites, robots must climb, descend, or traverse steep slopes. Future goals for exploration on other planetary bodies may require access to equally rugged terrain. 2.2 Implications In addition to furthering the development of a climbing robot for vertical natural terrain, the results of research in this area could provide fundamental insight into several related research areas. For example, this study could lead to the development of better strategies for robotic walking or dexterous manipulation. Human climbers often comment on an increase in balance and an expanded range of movement in everyday activity as they become more proficient at the sport. This enhanced mobility is often referred to as “discovering new degrees of freedom,” and is related to the idea of discovering useful new modes of mobility for ex-tremely complicated humanoid robots or digital actors. Also, the development of planning algorithms for climbing robots could lead to a better set of criteria for the design of these types of robots. These algorithms could be applied to candidate designs in simulation to determine the capabilities of the resulting robots, and thus to select a design. 3 Fundamental Issues There are five fundamental issues involved in climbing steep natural terrain: hardware design, control, sensing, grasping, and planning. A substantial amount of work needs to be done in each of these areas in order to develop a real climbing robot. This section describes the challenges involved in the first four of these areas; the planning problem will be discussed in more detail in Section 4. 3.1 Hardware Design A good hardware design can increase the perform-ance of the robot, and often can make each of the other fundamental issues easier to deal with. However, past use of hardware solutions in maintaining equilibrium generally resulted in a fundamental limitation on the terrain that could be traversed. Wheeled robotic systems have been used to ascend and traverse natural slopes of up to 50 degrees, to descend slopes of up to 75 degrees, and to climb over small obstacles in rough terrain. These systems either use some form of active or rocker-bogie suspension as in [12, 14-16], or use rappelling as in [1]. Similar results have been obtained using legged rappelling robots [3, 17] and a snake-like robot [4]. The terrain that these rovers can traverse robustly is impressive, but none of the existing systems has been shown to be capable of climbing natural slopes of 90 degrees or higher. Wheeled rovers and snake-like robots have an inherent grasping limitation that prevents their use in ascending sustained near-vertical or descending sustained past-vertical natural slopes. Existing legged robotic systems do not have this limitation, but still have bypassed the issue of main-taining contact with the slope by using rappel tethers. Reliance on these tethers prohibits initial cliff ascent, and limits the slope grade on cliff descent to below 90 degrees. A wide variety of robots capable of climbing vertical artificial surfaces is available. Most of these robots exploit some property of the surface for easy grasping. For example, some of these robots use suction cups or permanent magnets to avoid slipping [5-8]. Others take advantage of features such as balcony handrails [9] or poles [10]. However, the surface properties that are exploited by these robots generally are not available in natural terrain. In contrast, the simpler hardware designs used by [2, 11] had no such limitations. It is expected that solutions to the planning problem such as the one presented in this paper will allow basic natural vertical terrain to be climbed by similar systems, in addition to the ducts and pipes climbed by existing systems, and will suggest design modifications for better perform-ance. Future studies could address the use of other types of tools for grasping vertical natural surfaces, such as tools for drilling bolts or placing other types of gear in rock. The use of these tools would allow more challenging climbs to be accomplished, in the same way that “aid” helps human climbers [18, 19]. However, these tools bring an increase in weight and complexity, slowing movement and limiting potential applications. 3.2 Control There are three primary components of the control problem for a climbing robot: maintenance of equilib-rium, endpoint slip control, and endpoint force control. These three components are tightly related. In order to maintain balance, both the location of the center of mass of the robot and the forces from contacts with natural features must be controlled. Control of slip at these contacts is directly related to the direction and magnitude of the contact forces. Existing control techniques such as those based on the operational space formulation [20] could form a baseline approach to the design of a control architec-ture for a climbing robot. However, these techniques could be extended in a number of different ways to achieve better performance. For example, future research might address the design of an endpoint slip controller that is stable with respect to the curvature of a contact surface, rather than with respect to a point contact only. 3.3 Sensing For control and grasping, the robot must be capable of sensing the orientation of its body with respect to the gravity vector, the location of its center of mass, the relative location of contact surfaces from its limb endpoints, and the forces that it is exerting at contacts with natural features. For planning, the robot must additionally be able to locate new holds and generate a description of their properties, possibly requiring a measurement of levels of slip at contact points. Sensor integration, in order to acquire and use this information with algorithms for control, grasping, and planning, is a challenging problem. Existing engineering solutions are available which can lead to the development of a baseline approach in each case. For example, sensors such as those de-scribed in [21, 22] can provide basic endpoint force and slip measurements, an inertial unit and magnetic compass can provide position information, an on-board vision system can provide a rough characterization of hold locations and properties, and encoders can provide the location of the center of mass. However, the improvement of each of these sensors—in terms of performance, mass reduction, or cost reduc-tion—presents an open area for research. Although the performance of the planning frame-work that will be presented in Section 4 would be improved with better sensor information, it does not depend on a perfect model of the environment a priori. Since the framework leads to fast, online implementa-tion, plans can be updated to incorporate new sensor information as it becomes available. 3.4 Grasping The performance of a climbing robot is dependent on its ability to grasp “holds,” or features on a steep natural surface. It has already been noted that special-ized grasping schemes, relying on specific properties of the surface such as very smooth textures, pegs, or handles, cannot be used for grasping arbitrary natural features. The problems involved in grasping natural holds will be examined further in this section. Traditionally grasp research has been interested in either picking up an object or holding it immobile (also called “fixturing.”) Research in this subject dates as far back as 1876 it was shown that a planar object could be immobilized using a minimum of four frictionless point constraints [23]. Good overviews of more recent work can be found in [24, 25]. In this field an impor-tant concept is “force-closure,” defined as a grasp that “can resist all object motions provided that the end-effector can apply sufficiently large forces at the unilateral contacts.” [25] Nearly all research on grasps has focused on selecting, characterizing, and optimiz-ing grasps that have the property of force-closure. However, for the task of climbing a grasp need not achieve force-closure to be a useful grasp. For example, a robot may find a shelf-like hold very effective for pulling itself up, even though this grasp would be completely unable to resist forces exerted in other directions. For this reason, the techniques for  (a) (b) (c) (d) Fig. 2. Four different human climbing grasps, the (a) open grip, (b) crimp, (c) finger-lock, and (d) hand jam. selecting, characterizing, and optimizing grasps must be expanded significantly to apply to climbing robots. Characterization involves examining the direction and magnitudes of forces and torques (also called wrenches) that can be exerted by the grasp. For example, for one-finger grasps on point holds, an adequate representation of this information is a friction cone, which will be used for the planning algorithm described in Section 4. The idea of characterization also encompasses a “quality factor.” Measures of grasp quality have been researched extensively and are well reviewed in [26]. This work lists eight dexterity measures that include minimization of joint angle deviations and maximiza-tion of the smallest singular value of the grasp matrix. Other relevant research has been done using the concept of the wrench space. Using this concept, quality is defined as the largest wrench space ball that can fit within the unit grasp wrench space [27]. The volume of the grasp wrench space, or of more specialized task ellipsoids, could be used as a quality measure [28]. These ideas have been expanded to include limiting maximum contact force and applied in a grasp simulator to compute optimal grasps with various hands in 3D [29, 30]. However, the concept of grasp quality is ill defined for grasps that do not provide force-closure. Depend-ing on the direction that a climber wishes to go, different grasps may be of higher quality. Furthermore, grasp quality generally includes a concept of security or stability, and this too is ill defined for non-force- closure grasps. Again, depending on the direction of applied forces, the security of a grasp may change. The concept of hold quality must be defined before useful optimization is possible. Also, an efficient way of transmitting this information to a controller or planner is necessary to accomplish the climbing task. A qualitative classification of different types of grasps already exists in the literature for human climbers [19, 31]. In this classification, grasps are first broken into two categories, those meant for pockets, edges, and other imperfections on otherwise unbroken vertical rock faces, and those meant for sustained vertical cracks. Several examples of different face and crack grasps are shown in Figure 2. The literature gives a rough idea of the quality and use of each type of grasp in terms of criteria such as a perceived level of security, the amount of torque that can be exerted on a hold, and the amount of friction at the “power point.” Not only is this expert intuition qualitative, but also it is clear that human climbers need to perform additional grasp planning for specific cases. As put by Long, “There are as many different kinds of holds as there are ways to grab them [31].” However, this intuition can be used as a starting point for determining meaningful quantitative criteria for grasp selection and optimization. A comparison of the climbing literature with past work on robotic grasp planning reveals several other fundamental differences between the two applications that may become important in future research. For example, many climbing holds are very small, so the fingers used in a climbing grasp often have large diameters relative to the object to be grasped. Litera-ture on robotic grasping primarily considers the case where the fingers have small diameters relative to the object. In addition, some climbing grasps, as men-tioned above and shown in Figure 2, are based on jamming fingers in a crack. This technique is very different from one a robot might use to pick up an object, and requires a high degree of flexibility and small degrees-of-freedom in order to “un-jam” the fingers. Clearly, continued work on climbing robots eventually will lead to the consideration of a wealth of new issues in grasping. 4 Planning The planning problem is the fifth fundamental challenge for climbing robots in natural terrain. Details of the motion-planning framework presented in this section are given in [32]. 4.1 Challenges The planning problem for a climbing robot consists of generating a trajectory that moves the robot through a vertical environment while maintaining equilibrium. This problem is challenging even for human climb-ers! Climbing is described by Long as a “singular  (a) (b) (c) Fig. 3. Three different human climbing “moves,” the (a) back-step, (b) stem, and (c) high-step. challenge, where each ‘route’ up the rock is a mental and physical problem-solving design whose sequence and solution are unique. Every climb is different [31].” Much of the sequence for a particular route might be composed of one of a variety of different types of “moves,” such as a back-step, stem, mantel, high-step, counterbalance, counterforce, lie-back, down-pressure, or under-cling. Some of these moves are shown in Figure 3. Each “move” is a learned technique for maintaining balance that may seem counterintuitive. In addition to these heuristics, movement through a large number of other very specific body positions might be necessary to progress towards the top of a climb. The importance of planning a sequence of moves before actually climbing is emphasized by Graydon and Hanson [19], who recommend that climbers “identify and examine difficult sections before [they] get to them, make a plan, and then move through them quickly.” The human motivation for this approach is primarily to minimize the effort required for each move and to conserve energy, since most people have hard strength and endurance limits. The planning problem for a climbing robot is quite similar. The robot likely will be equipped with actuators that can exert high torques only for short amounts of time, so planning a sequence of moves before climbing is important for a robotic system as well. Likewise, a climbing robot will be subject to the same hard equilibrium constraints, and will need to select between a similarly wide range of possible motions. Therefore, the development of a planning algorithm for an autonomous climbing robot is a very challenging problem. 4.2 Related Work The search space for a climbing robot is a hybrid space, involving both continuous and discrete actions. Many different methods are available for motion planning through continuous spaces, including cell decomposition, potential field, and roadmap algo-rithms [33]. Discrete actions can be included in these methods directly, for example at the level of node expansion in roadmap algorithms, but this approach generally leads to a slow implementation that is specific to a particular system. Previous work on motion planning for legged robots has developed tools for addressing these hybrid search spaces for some systems. This work can be categorized by whether or not the planning is done offline, in order to generate a reactive gait, or online, in order to allow non-gaited motion specific to a sensed environment. Gaited planners generate a predefined walking pattern offline, assuming a fairly regular environment. This pattern is used with a set of heuristics or behav-iors to control the robot online based on current sensor input. Gaited planning was used by [2, 11], for example, to design patterns for climbing pipes and ducts. Other methods such as [34] are based on the notion of support triangles for maintaining equilib-rium. Stability criteria such as the zero-moment-point have been used to design optimal walking gaits [35]. Dynamic gaiting and bounding also have been demonstrated [36-38]. Recent work [39, 40] has attempted to provide unifying mathematical tools for gait generation. Eac