Ambulating Robots for Exploration in Rough Terrain on ... - CiteSeerX

Ambulating Robots for Exploration in Rough Terrain on Future Extraterrestial Missions Dirk Spenneberg, Frank Kirchner and Jose de Gea Robotics Group Department for Mathematics and Computer Science University of Bremen Bibliotheksstr. 1 28359 Bremen, Germany {spenneberg, kirchner, jdegea}@informatik.uni-bremen.de Abstract We present the SCORPION project and the new ESTEC funded ARAMIES project to design, built and program a multi-functional, multi-degree of freedom, ambulating robot that is capable of autonomous operation in extremely difficult environments. The special focus in the ARAMIES project is on very uneven and steep terrain, e.g. crater slopes. The software approach of both robots is based on a bio-inspired behavior based control concept, which produces very robust and stable walking patterns.

I. I NTRODUCTION Extraterrestrial sites with high scientific potential prove often to be extremely difficult to access. Such environments comprise: • extremely difficult (rocky, sandy) substrates • steep terrain (canyon walls, cliffs) • uneven, hashed terrain Conventional wheeled rovers are not able to provide access to these sites. Thus new technology has to be developed. Because of the impressive capabilities of walking animals in natural and rough environments, it is likely that biomimetic ambulating, autonomous robots can provide the necessary mobility. Therefore they are perfect vehicles for the scientific investigation and exploration of extraterrestrial sites. The ARAMIES Project (ESTEC Contract: 18116/04/NL/PA) deals with the development of such a system. Because of the targeted weight and size of the proposed robot, it will be possible, but is only partially subject of this research project, to deploy multiple of the proposed devices. This would further increase the potentials for scientific studies in three general aspects: • parallel investigations at different locations, • joined, team oriented exploration / investigation with different sensor packages at the same site, • team oriented approaches for access to extreme sites such as canyon walls and steep cliffs. Therefore we see the relevance of the proposed research in the overall enhancement of scientific investigations with respect to both, qualitative and quantitative aspects. This project is based on experiences we gained with the eightlegged SCORPION robot. At lot of the lessons we learned with the SCORPION will be helpful in the ARAMIES project. Furthermore the performance of the SCORPION gives already an idea of what we have to expect from the new robot. II. T HE SCORPION ROBOT The SCORPION project1 aimed at the development of biomimetic walking robots for rough terrain. Possible future fields of application for robots based on legged locomotion are the work in dangerous, highly unstructured, rough and unpredictable environments, where mobility is critical, for example search and rescue missions in collapsed buildings. The mobility of present wheeled and also tracked vehicles is too limited for such tasks. Therefore it is crucial to develop new vehicles, which are based on new locomotion technology. The developed SCORPION robot has eight legs and measures 65cm from front to back. The width depends on the posture of the legs and varies between 20cm and 60cm. In its typical energy-saving M-shape walking position the robot is 40cm wide. It weights 11.5 kg including batteries. Each leg has three DOF: a thoracic joint for protraction and retraction, a basal joint for elevation and depression and a distal joint for extension and flexion of the leg. Thus the whole system comprises 24 joints. The joints are actuated by 6 Watt DC-motors which drive planetary gears. The legs feature a spring element in the distal segment to reduce the mechanical stress and for measuring the ground contact force via an in parallel integrated linear potentiometer. The robot acquires the following proprioceptive information: the position of each joint, the amount of current drawn 1 sponsored

by DARPA (Grant No. N0014-99-1-0483

Fig. 1.

The SCORPION Robot in an Outdoor Test

by each motor, the tilt in 2 dimensions (pitch, roll) and the load on each foot tip. Furthermore it is equipped with a compass and a forward facing infrared distance sensor. If the robot is not used in a full autonomous mode, it is possible to teleoperate the system. The system has a bi-directional communication link for video transmission with a PAL CCD camera and for data exchange between the robot and an external laptop with a communication and control interface. The control hardware is onboard and features a Motorola MPC555 microcontroller and an XILINX Virtex 400 E FPGA. The MPC555 is used for the processing of the non-leg sensors and for the behavioral control of the system. The FPGA is used for the processing of all leg-sensors (position and drawn current of each joint, foot-tip) and for the 20Khz PID-control of the joints. III. B IO -I NSPIRED C ONTROL To use a standard kinematical model, as used for industrial robotic arms, for describing the dynamics of such a system is very difficult. For a system with n degrees of freedom one gets at least n coupled equations. Furthermore the equation a non-linear due to the rotary joints. Computations with such a set of nonlinear equations are non-trivial, which makes the use of an detailed kinematical model for the control of autonomous walking agents very unattractive. However there are some approaches to simplify the equations, e.g. the Free Body Diagram Approach[1]. In contrast biological systems like spiders or insects are very well able to solve this complex nonlinear problem. Given the fact that they are using in comparison to current computers a comparatively slow computing nervous system, there must be other solutions to the problem of robust walking through rough terrain. Biologist have proposed different models in dependency of their experimentally gained data and observed animals. All these models have in common that they rely heavily on decentralized control. They can be divided in two major assumptions. The first assumption is that motor control is based on so called ”Central Pattern Generators” (CPG) wherein the actuators are controlled by an endogen pattern produced by a central neuronal oscillatory mechanism. The second assumption is reflex-driven control where, in contrast to the first one, the state of the actuator is only a function of the interaction with its environment. The Walknet [2] is a pure example of the later

Fig. 2.

Scheme of the Motion Pattern Control Approach

idea. Here the movement of neither a single leg nor the coordination of all legs is centrally pre-programmed, thus instead of using one or more central controllers with global knowledge, each leg uses its own controller with only procedural knowledge for the generation of the movement of the leg. Hence, production of the gait is an emergent property of the whole system, in which each of the six single-leg-controllers obeys a few simple and local rules in processing state-dependent information about neighboring legs. Parts of this control approach were successfully tested on the six-legged robot TARRY II [3]. In contrast the control architecture developed by Prof. Joseph Ayers for control of the Lobster robot [4] is a pure example for the CPG approach. Here the rhythmic motions of lateral left, lateral right, forward and backward walking are implemented as pre-programmed finite state machines. The control to execute one of these finite state machines is done by a command neuron like structure [5]. Both approaches have in common that they need only a low amount of computational power and thus are suitable for autonomous walking robots. A pure reflex based control acts only on the basis of sensory input, whereas the CPG-based control is able to produce rhythmic motion without the need of sensory feedback. Thus in case of false, unreliable or insufficient sensor data (e.g. at high speeds) the CPG-approach is more robust. On the other hand in a very unstructured and dynamically changing environment the endogen produced motion pattern might be highly inadequate. Therefore the control approach (see figure 2) we have developed for the SCORPION robot aims at using the best of both worlds. The idea is to have rhythmic motion behaviors (RMBs), which can be activated with different strengths by the central behavior-based control level. The RMBs control the motion of the system like a CPG-based system when the disturbances from the environment are rather small. These behaviors simultaneously influence the amplitude and the frequency of the thoracic, basal and distal oscillators (OST,OSB and OSD). The oscillators are connected to a common clock which is used for local and global (in relation to the other legs) synchronization purposes. The oscillators output is a rhythmic signal which can be described with splined sinusoid waves. It describes the trajectory of the corresponding joint in its angle space. Thus it represents the desired motion. This angle position signal is translated via the motoric end path into pulse width modulated (PWM) signals to drive the motors. Inline with this rhythmic input to the motoric end path are a set of perturbation specific reflexes, which are implemented as ’watchdogs’. If greater disturbances are sensed these reflexes (e.g. the ”Stumbling Correction reflex” [6]) are triggered and override the signals of the oscillators with precompiled motion signals to stabilize the system. The RMBs enable the system to move forward, backward and lateral and to turn with different radii. By activating more than one of these the RMBs’ effects can be combined through a overlaying process. For example, if the ’move forward’ and the ’move lateral left’ RMBs are activated at the same strength the resulting motion is a diagonal forward motion to the left. The overlaying process ensures that the transition from one motion to another is a smooth and quick motion, thus ensuring that the system keeps stable. An advantage of this is that the change of the motion direction doesn’t require to stop the robot first.

In addition to the RMBs the architecture provides the higher behavioral level with so called Posture Motion Behaviors (PMB) as means of control of the posture of every leg. For example the height and the tilt control behaviors on the central control level are using these PMBs to stabilize the system while it is walking. Again the overlay process combines the influence of the PMBs with the influence of the RMBs on the actuators. Thus it is possible to change the height of the main body, while walking, by just outstretching the legs. All these mechanisms together enable the system to walk at quite constant speeds through rough terrain. Thereby it is possible to walk with the same software architecture over a high variety of different substrates like rocks, sand, mud, grass, concrete and asphalt. Its maximum speed over flat terrain is half of a body length (30cm/sec). The system is able to overcome obstacles as high as a full outstretched SCORPION and it can climb up ramps up to 35% and still overcome small obstacles, like 8cm high pipes. Furthermore the additional postural control, gives the user in addition to the rhythmic locomotion the exact control over every joint, which can be used for deliberative control. This can be used for a wide variety of interesting future behaviors, for example the posture of a single leg can be changed in the way that it can be used to carry simple objects on the back of the robot. We implemented and successfully tested a behavior to grab a beam with one leg and then walk away with it on the remaining seven legs. Other experiments have shown, that is it possible by only changing the posture of the system and using the unchanged RMB for forward walking, up-side down brachiation along a beam can be executed. The only difference between brachiation and forward walking in our approach is the activity level of the forward RMB on the legs and the posture of all legs. So the locomotion control level provides the programmer with a simple but very powerful interface for locomotion control. At the moment we are in the progress to enhance this locomotion control with learning components, which we also want to use in the ARAMIES project. The biomimetic approach enables us to investigate the possibilities of using proprioceptive data to generate information about the environment by interacting with it in a non-trivial way. When interacting with the environment, e.g. walking over it, the system produces a multitude of proprioceptive data. In a first experiment we started to explore the possibilities to use this data to detect on which substrate the robot is moving. This would be a very interesting new feature for walking robots, which makes especially sense in almost unknown environments or where at least the properties of the ground are unknown, for example on extra-terrestrial missions. We have build an indoor proving ground, which contains different substrates like sand, rocks and gravel. Especially the difference between the sand and gravel is not very large, which makes this proving ground an interesting challenge for spatial categorization apporaches. To classify the proprioceptive sensor data we gain, when the robot walks in this proving ground, a supervised learning approach based on growing cell structures [7] is used. This first experiment was very promising and the robot was able to classify 65% of its environment correctly by only using proprioceptive data. IV. T HE ARAMIES PROJECT Based on the experiences with the SCORPION robot we have started to develop a more flexible robot for rough and steep terrain. The aim in the ARAMIES is to develop a walking system for extremely difficult terrain, especially steep and uneven slopes. It is planned, that the robot is able to carry different science packages (mission specific sensor equipment) for data acquisition, analysis and/or sample return. To have maximum flexibility in the hardware of the robot, we decided to focus in the first year of the project on the development of very modular control hardware and mechanics. In the software part we will focus more on extending our biomimetic approach to develop a framework architecture, that allows us to integrate bio inspired, low level mobility control schemes with medium level, reactive controllers and high level, deliberative, planning and scheduling algorithms. The project started 6 months ago. The following subsections will describe the hiherto results. A. The Electronics In the field of electronic design we faced the challenge to design a small sized, low energy and highly integrated electronics package that provides robust and fault tolerant operation under extreme conditions and which can be easily adapted to the computational demands of future software, e.g. mission planer. Therefore a new control electronics was build, which was based on experiences we gained with the SCORPION control board. The aim was to develop a very modular system regarding its physical installation and possible upgrading.

Fig. 3.

The Carrier Board Hosting the MPC565 Phytec Board and the XILINX Virtex E 600 Board

The electronics is composed of three main boards (which form the so-called Bremen Board platform) and two optional / external ones: • The microcontroller board. The core of the system is a phy-CORE MPC565 board. This is an insert-ready Single Board Computer populated with Motorola’s PowerPC MPC565 microcontroller and external FLASH-ROM (4Mb) and SRAM (16Mb). The MPC565 microcontroller is the next generation of the MPC555 microcontroller, which was used in the SCORPION. • The FPGA Board. It includes a XCV600E FPGA from Xilinx. This FPGA has nearly 1 million gates and, as an important feature, it supports differential signalling that allows us to use the LVDS bus to connect the motors to the FPGA thus minimising the cabling. This board is responsible for acquiring information from the motor sensors and driving them with the fastest response to input changes. In this particular case, it implements an Address Decoder, a Dual-Port RAM, 8 ADC modules to read the information coming from the Motor Boards, PID controllers and PWM modules for each of the 32 DC Motors. • The Carrier Board. This board serves as a backplane board, where both microcontroller board and FPGA board are plugged in. Primarily, the Carrier Board has to act as a main board where the other ones will be plugged in. However, it will be not only a ”passive” board without functionality. In fact, it includes important features: different on-board sensors (accelerometer, gyroscope..), Dual-Port RAM, PC-104 interface, LVDS drivers, signal conditioning circuits for infrared and ultrasound sensors and the necessary drivers to interface with other boards. • The Motor-Sensor Board. This board is the only one that is not directly plugged to the Carrier Board. In this case, and since the motor boards will be split along the robot, the connections to the Carrier Board will be made using two pairs of differential signals (LVDS bus). Each board is capable of driving 4 DC Motors of up to 5A each. Furthermore, it is able to read the following information from the motors: position, current, encoder information and pressure, force or torque signals. The microcontroller board supports up to 8 of these Motor/Sensor Boards, since it has to be possible to manage up to 32 DC motors at a time. The LVDS bus is essentially a signaling method used for high-speed transmission of binary data over copper. It uses a lower voltage swing than other transmission standards. This low voltage differential is what delivers higher data transmission speeds and inherently greater bandwidth at lower power consumption. LVDS can achieve signaling rates as high as some Gigabits per second (Gbps). It consumes as little as one-eighth the power of RS-422 drivers. Since the receivers respond only to differential voltages, they are relatively immune to noise such as common-mode signal reflections. In addition, LVDS emits less electromagnetic interference (EMI) than other data transmission standards. The dimension of a motordriver board are 100 x 40 mm. • The PC-104 Board is an extension board that allows us to add a full industrial PC104 computer, which can

communicate with the MPC565 via a Dualport-RAM, which is accessed through the ISA-Bus. By making use of the PC-104 bus we can also add future capabilities to the system, for example GPS, Ethernet or GSM. As stated, the electronic design will have a main core, the PowerPC microcontroller. But strictly speaking, we should talk of a second core, the FPGA. Although the latter one needs information from the first, we could see them as parallel processors working on different tasks at the same time but sharing information. At a low level, the Motor Boards will read information from the motors (position, current, pressure, torques..). This information will travel along the LVDS drivers to the FPGA inputs. There, according to a re-programmable controller, the necessary PWM signals to drive the motors will be generated. These signals will travel again out of the FPGA board through the LVDS bus and will get to the power converters present on the Motor Boards, thus driving the motors with the desired pattern. Since no software in included in this loop but the hardware implementation on the FPGA, a very fast response to events will be achieved. At a higher level, the microcontroller will process information coming especially from the exterioceptive or higher sensors of the robot (infrared, ultrasound, compass, accelerometer). This information will be processed on the behavioral level odf control to modulate the low-level and local locomotion control of the joints, thus changing the parameters of the motor controller on the FPGA. In principle we are using here an enhanced a more modular SCORPION board. So we were able to keep and extend all the exiciting features of the SCORPION board, while making the hardware much more flexible in terms of physical installation and extension possibilities. B. The Mechanical Design The Mechanical Design was aimed at maximum flexibility and light weight design of the mechanical components while maintaining stiffness, robustness and dynamic stability. Flexibility is one of the key-issues, when building kinematical complex systems, because it allows for rapid changes in the morphology of the robot. In our project it is yet unclear, what kinematical configuration will turn out to be the most suited for the task of climbing in steep slopes, therefore a mechanical concept which allows to test different configurations is desirable. .

Fig. 4.

A first Leg Prototype

Fig. 5.

A Four-legged Configuration

Hence we developed a construction kit to be able to rapidly change the kinematics of our robot. The kit consists of basic elements, which can be connected to each other in a modular way. It is possible to build different combinations of joints and limbs and thus different extremities, which can have a various number of active joints. The mechanical concept for the ”Extremities” is based on single, motor-driven rotating pipes. These modular segments can be easily combined to a mechanical chain (an Extremity), which can be as complex as necessary. For

an example see the first prototype leg in figure 4. For the end effectors of these Extremities we have foreseen a claw like device, which also will able to grip objects. These end effectors will be used on one hand as feet and have therefore also damping characteristics to compensate for mechanical stress and on the other hand as manipulators. The claws are still in development, till then we are using a preliminary static foo, which you see in fig. 4. Figure 5 illustrates how the different elements of the construction kit can be combined to a full robot. V. O UTLOOK We plan to have a first integration study ready using the new construction kit and the new electronics by April 05. This integration study will already demonstrate basic walking capabilities and is the major milestone of the first year of the project. The main scientific/technical goal of this three year project will be the presentation / evaluation of the physical robot in a simulated exploration / investigation mission under earth conditions. The system will be completely self-contained in terms of actuation, sensing, computation, communication and energy.

Fig. 6. Left Picture:The idea of an Cooperative Exploration with a SCORPION like Robot as a Scout Robot, which is Launched from a Rover Using a Tether System. Right Picture: The Same descend from the Rover Perspective.

In 2006 in the final stage of testing we want to deploy the system (including a camera based sensor package) to explore a steep cliff. This idea is depicted in the figure 6. This has some similarities with the DANTE II Mission [8], but instead of making scientific research on the crater bottom we are interested in the different possibilites to do research in the steep face of the crater. In this test the system will be attached to a second robotic system (which will be simulated by a fixed operator desk instead of a second robot, as a second hardware robotic system is beyond the scope of this project) via the tether deployment system. R EFERENCES [1] J. Barreto, A. Trigo, P. Menzes, J. Dias, and A. D. Almeida, “Fbd - the free body diagram method, kinematic and dynamic modelling of a six leg robot,” in Proceedings of AMC’98 - 5th International Workshop on Advanced Motion Control, Coimbra, Portugal, 1998. [2] H. Cruse, J. Dean, T. Kindermann, J. Schmitz, and M. Schumm, “Walknet - a decentralized architecture for the control of walking behavior based on insect studies,” in Hybrid Information Processing in Adaptive Autonomous Vehicles, G. Palm, Ed. Springer, 1999. [3] M. Frik, M. Guddat, D. Losch, and M. Karatas, “Terrain adaptive control of the walking machine tarry ii,” in Proceedings of the European Mechanics Colloquim, Euromech 375, 1998, pp. 108–115. [4] J. Ayers, “A conservative biomimetic control architecture for autonomous underwater robots,” in Neurotechnology for Biomimetic Robots, Ayers, Davis, and Rudolph, Eds. MIT Press, 2002, pp. 241–260. [5] I. Kupferman and K. R. Weiss, “The command neuron concept,” Behav. Brain Sci., vol. 1, pp. 3–39, 1978. [6] H. Forssberg, “Stumbling corrective reaction: A phase-dependant compensatory reaction during locomotion,” Journal of Neurophysiology, vol. 42, no. 4, July 1979. [7] B. Fritzke, “Growing cell structures - a self-organizing network for unsupervised and supervised learning,” Neural networks, vol. 7, no. 9, pp. 1441–1460, 1994. [8] J. E. Bares and D. Wettergreen, “Dante ii: Technical description, results, and lessons learned,” Internat. Jornal of Robotics Research, vol. 18, no. 7, 1999.