Design and control of a small size humanoid robot

Chapter 1 Introduction It has always been the dreams of many for man to co-exist with humanoid robots, to live and work in the same environment. Japan, as a leading country in robots and their applications, has incorporated robotics in their manufacturing industries for years. However, most of the robots involved are limited to robot arms that are fixed to the ground and tasks allocated to them are straightforward and repetitive. The desire to build robots resembling ourselves is reflected in the works of many researchers in recent years, where a significant focus is placed on building humanoid robots. Robotics competitions around the world have also included humanoid category in recent years, and it is perceived as one of the most challenging groups that would draw crowds of spectators. Team RO-PE, formed in 2002, by the Mechanical Engineering Department of National University of Singapore, has also been playing a part in striving to advance the technology in humanoid robots and it had participated in the humanoid category of both RoboCup and FIRA, two famous CHAPTER 1: INTRODUCTION 2 international robotics competitions. RO-PE-V, the fifth humanoid robots built by Team RO-PE, had represented the team to participate in RoboCup 2006 and 2007, and had achieved encouraging results. And RO-PE-V is employed as the subject of this thesis. This project involved the design and building of a humanoid robot, RO-PE-V. With RO-PE-V setup as a platform, walking control had been implemented. Experiments on localization and slope walking were also performed. And these are presented in this thesis in 8 chapters. Chapter 1 Introduction – Some background information on the topic are provided in this chapter, the scope of the thesis is also laid down. Chapter 2 Literature Review – In this chapter, related works from other researchers are discussed, reviewing the current state of technology and general approach in this field. Chapter 3 Sensors, Actuators and Computer Systems – The important hardware mounted on the robot are explained in this chapter. Chapter 4 Mechanical Design – The design philosophy and approach are presented in this chapter. Chapter 5 Walking Control – In this chapter, the approach used to control the walking of the robot is presented. CHAPTER 1: INTRODUCTION Chapter 6 3 Slope Walking – Some experiments are done on a simple approach to slope walking, the logic of this approach will be presented in the chapter. Chapter 7 Localization – Localization in a colour-coded environment (RoboCup Competition) is experimented with the robot as the platform and will be discussed in this chapter. Chapter 8 Conclusions and Recommendations – In this chapter, conclusions to this project and this thesis are given, some recommendations for further investigation in this topic are also provided. Chapter 2 Literature Review The study of humanoid is an interesting field of research which is highly complex and multi-disciplinary. It had for a long time been only the dream and fantasy of man, and exists only in science fictions, novels or movies. The earliest engineering records of humanoid would probably be the design of a humanoid automaton by Leonardo da Vinci in around year 1495 [30], and it is still unknown whether it was physically built or just a paper design. This line of research remained largely unexplored for many years until the last three to four decades. This is primarily due to the fact that technology at that time, especially in terms of hardware, was still unable to handle the stringent requirements of humanoid robot, making the topic extremely difficult to handle. There is a significant advancement in humanoid research in the last three to four decades. Waseda University from Japan began their humanoid research in about 1966 and built the world first full-scale humanoid, WABOT-1, in 1973 [1][30]. The interest in humanoid research did not stop at research institutes and universities and commercial companies also took up the challenge in research. The pioneers in this field is the Japanese car manufacturing giant, Honda, which began their research in about 1986. Many versions of humanoid robots had evolved from CHAPTER 2: LITERATURE REVIEW 5 Honda through the years, with ASIMO being its latest version [5][28]. HRP-2 is another famous humanoid produced by Kawada Industries Inc. [2][3], and it is able to cooperate with human to carry some load. Furthermore, it had demonstrated the ability to get up from a face-down position, which is very challenging given its height of 158cm. The three robots mentioned above are relatively larger robots that have heights of more than 1m. They are expensive and more difficult and dangerous to handle. Many researchers then turn to scaled-down humanoids, of height of about 50cm, where in terms hardware, are much more affordable. Qrio from Sony [4] and HOAP from Fujitsu [40] are two commercial small size humanoids that are produced a few years back. Though the robots could be for sale, the price is extremely steep for them to dominate the small size humanoid market. Realizing the growing interest in humanoid robots, motors manufacturers are coming up with their own humanoid for sale. URIA from Robotis [32], ROBONOVA from Hitec [33] and KHR-1HV from Kondo [34] are some of the relatively low priced humanoid available in the market. RoboSapien [35] is another budget humanoid built for the toy industry. Because it is meant to be a simple toy, it does not carry a powerful processor that would make it more ‘intelligent’. In fact, some researchers use RoboSapien as a walking platform, replacing the processor with a more powerful one like a PDA for more intensive computation like image processing and collaboration between robots [24]. International robotics competitions like RoboCup [36] and FIRA [37] had also called upon researchers from around the world to construct their own humanoid CHAPTER 2: LITERATURE REVIEW 6 robots. Among them are VisiON NEXTA from Vstone Corporation [38] and NimbRo from University of Freiburg [39], which had both shown exceptional performance in competitions. Fig. 2.1 shows the different humanoids seen around the world today. (a) WABOT-1 [30] (d) QRIO [42] (b) ASIMO [28] (e) HOAP2 [40] (c) HRP-2 [41] (f) URIA [32] CHAPTER 2: LITERATURE REVIEW (g) ROBONOVA [33] (i) RoboSapien [35] 7 (h) KHR-1HV [34] (j) VisiON NEXTA [38] (k) Nimbro [39] Fig. 2.1 Humanoid robots constructed for different purposes. Though there are already many humanoids built and walking, there are still many areas in this topics that are not fully covered, and it would still take a lot of effort and time before these robots could be made to work safely (for both human and the robots) in an unstructured area. The current research approach in humanoid could broadly be classified into three areas, (1) mechanical design and hardware, (2) walking control and (3) artificial intelligence. CHAPTER 2: LITERATURE REVIEW 2.1 8 Mechanical Design and Hardware In the area of mechanical design, one of the important areas is to decide the number and locations of degrees of freedom for the biped. Fred R. Sias, Jr and Yuan F. Zheng [9] had done an indepth research on this and came to a conclusion that eight degrees of freedom are required on each leg to have a good approximation of human gaits by a biped robot. However, they also remarked that the degree of freedom with a vertical axis at the ankle is unnecessary for most gaits used for locomotion, while the degree of freedom at the foot is significant only for rapid walking. Therefore, in most situations, a leg with six degrees of freedom (three at the hip, one at the knee and two at the ankle) is employed so as not to complicate the design of the robot. P2 from Honda is example of humanoid with six degrees of freedom on each leg [5]. Valuable design experience and lessons learnt are shared among the research community through publication. Research work by Honda [5] shows that impact absorption at the foot is of paramount importance. Not only that it would help to protect the hardware on the robot from potential damage caused by the impact force, damping by rubber-like protection could also help to prevent vibration by acting as a mechanical lowpass filter. It was pointed by the designers of SDR-4X from Sony [4] that the yaw axis of the leg should be offset towards the back. By doing so, a wider turning angle could be achieved by this yaw motion before having the two feet hitting each other. Fig. 2.2 explains the logic of the shift in a pictorial form. CHAPTER 2: LITERATURE REVIEW 9 Fig. 2.2 Offsetting the yaw axis to achieve wider turn angle [4]. Motors and power transmission are necessary components of humanoids and some robot researchers like Honda and Sony are using their own customized motors for actuation. While harmonic gears are getting popular in the large humanoid robots community, normal gearbox remains the common selection by small size humanoid robots as they are usually integrated with motors as a compact package by the manufacturer and are much cheaper. However, backlash would be a potential problem for using gearbox, compromising precision in motor control. Timing belt is an alternative to overcome the problem of backlash in gearbox system, HOAP2 from Fujitsu uses timing belt for power transmission. As for the main frame of the robot, the general idea would be to have the building material to be as light and as strong as possible. However, one would expect a strong material to be heavy and a light material to be weak. Therefore, compromise on this is required to identify an optimum material. Common choice for this application would be aluminum alloy, well known for its low density of about 2700Kg/m3, and also machinability. Recently, there is a trend for small size humanoid robots to use composite materials like carbon fibre sheets or tubes, CHAPTER 2: LITERATURE REVIEW 10 which has a typical density of about 1750Kg/m3, as the structural material. NimbRo from the University of Freiburg is an example of humanoid robot built with carbon fibre. 2.2 Walking Control Given the complexity of a humanoid robot, walking stably is a challenging task. Even human beings need months to learn how to walk. Bipedal walking control has been the focus for many researchers. And many approaches to achieve stable walking had been considered, and they could generally be classified into five main categories [12], (1) model-based, (2) ZMP (zero moment point)-based, (3) biologically inspired, (4) learning and (5) divide-and-conquer. 2.2.1 Model-based approach In model-based approach, mathematical models derived from laws of physics are used to generate control algorithm. Approximations are made to vary the complexity of the mathematical model. Using this approach, Kajita et al. [14] had come out with the linear inverted pendulum model by approximating that the mass of robot legs to be negligible compared to the body mass. The system would then be similar to an inverted pendulum pivoted at the ankle joint. By constraining the mass to move in a linear path, a closed-form solution could be found for the linear differential equation. CHAPTER 2: LITERATURE REVIEW 2.2.2 11 ZMP-based approach ZMP or zero moment point is a widely adopted concept in humanoid robotics. The term was first coined by Vukobratovic [27], which refers to the point on the ground where the resultant of the reaction forces from the ground acts on the robot. It is believed that ZMP is an indication to the stability of the walking biped. Therefore, by planning the desired ZMP positions, the required positions of the centre of mass could be obtained, and through inverse kinematics, obtaining the joint trajectories. Wasaeda University was the first to implement this control approach on a real robot [30]. 2.2.3 Biologically inspired Passive dynamic walking is a form of walking behaviour that was discovered by Tad McGeer [6]. It was shown that a passive walker could walk down a slope based on just gravity and no actuation was needed. This was also inspired by the fact that human being does not need to exert a lot in order to walk. 2.2.4 Learning Learning is a natural concept for walking control for the fact that human beings need to learn in order to walk properly. The general idea in learning is to allow the robot to try to walk and gain experience through the process, repeating and improving the task until the final goal is achieved. 2.2.5 Divide-and-conquer As the name suggest, this is a very common approach to a complex problem, where this complex problem is handled by breaking down to a few simpler sub- CHAPTER 2: LITERATURE REVIEW 12 problems and be tackled individually. In the case of bipedal walking, it could be broken down into the frontal and sagittal plane for better analysis. 2.3 Artificial Intelligence In the field of humanoid research, works done on artificial intelligence are rather limited. A typical interpretation on artificial intelligence for humanoid would be for humanoid robots to perceive the environment and make appropriate decisions, it is also suppose to learn and become more intelligent in the process of learning. This would then depend on the task allocated to the robot, and currently, the tasks given to humanoid robots are rather simple and they usually operate in a structured area. ASIMO from Honda [28] had demonstrated an encouraging level of intelligence by recognizing voice of people and moving around with people. But still, there are much works to be done before humanoids could really be intelligent. Chapter 3 Sensors, Actuators and Computer Systems A fully autonomous humanoid robot could complete a task by itself, without assistance from outside system. Thus, all the components that are required to complete the task has to be mounted on the robot. The three groups of basic components that are needed are (1) Sensors, which receive signals from the surrounding environment and feedback to the brain of the robot for necessary reaction, (2) Actuators, which are the components that perform the physical movements upon receiving signals from the controller, and it is the combination of motions from many actuators that allows the humanoid robot to walk and (3) Computer Systems, which acts as the brain of the robot, take in signals from sensors, compute, come out with the necessary reaction plan and produce the signal to instruct the robot or, the actuators, the actions to be carried out. CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 3.1 14 Sensors For a robot to be fully autonomous, it is necessary that it carries some form of sensors that it could use to collect information on the surrounding environment and also on its own status. And with these information, the robot could come out with the necessary reaction plan for execution later. For human, we are equipped with numerous sensors. Our eyes, ears, nose, tongue and skins are the most basic sensors everyone is familiar with. So it would be intuitive that the primary sensor of RO-PE-V to be its vision system. 3.1.1 Vision System An omni-directional vision system was selected to be the primary sensor of ROPE-V, instead of the conventional pan-tilt vision system. The concept of omnidirectional vision system was first proposed in 1970. The idea is to have a camera looking up at the curved mirror that is reflecting the image of the surrounding. Fig. 3.1 illustrates the schematics of this vision system. Light rays Fig. 3.1 Schematics of omni-directional vision system. CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 15 The main advantage of using the omni-directional vision system is that it allows the robot to see 360o around itself, identifying several landmarks simultaneously and this feature is especially useful for localization which will be discussed in Chapter 7 of this thesis. Another reason for using this new type of vision system is weight reduction. For the conventional pan-tilt system, actuators are required to execute the pan and tilt motions for the camera to see a larger region. But since the omni-directional vision system is already able cover 360o, there is no need for the pan and tilt motions, thus, shaving the weight of two actuators that would about 50g each, whereas the additional mirror in the omni-directional vision system only weighs about 30g. However, these advantages are accompanied by some short-comings of the system. The main disadvantage of an omni-directional vision system is that there will be distortion in the image captured by the camera due to the fact that the camera is seeing the surrounding through a curved mirror. With this, the distance of an object could not be obtained straight-forwardly, the distorted image also affects the visibility of objects that are relatively far away. Fig.3.2 shows an image obtained through the omni-directional vision system. In addition, there are also blind spots for this vision system, the view of regions just around the robot are blocked by the shoulder and the body of the robot, though this could be overcome by some actions of the robot to clear the obstructions. CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 16 Fig. 3.2 An image captured by the vision system of RO-PE-V. 3.1.2 Magnetic Tilt Switch Two magnetic tilt switches from Assemtech are mounted on RO-PE-V to detect the orientation of the robot with respect to the ground. They are important sensors because they provide the feedback on the status of the robot, i.e. whether the robot has fallen down, and the appropriate reactions could be carried out, for example, the ‘getting up’ routine. Fig. 3.3 shows the picture of the tilt switch employed. Fig. 3.3 Magnetic tilt switch (MTA 240) from Assemtech. The magnetic tilt switch is really an on/off switch governed by the position of a movable ball bearing, which rolls along a guided path depending on the orientation CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 17 of the tilt switch. These tilt switches, thus, provide digital signals to the computer system of RO-PE-V for decision making. Fig. 3.4 shows the schematics of the working principles of the magnetic tilt switch. When tilted, magnet falls to the right After the magnet falls to the right, the metal connectors will be attracted towards, the magnet, closing the circuit in the process. Fig. 3.4 Schematics of the working principle of the magnetic tilt switch. 3.1.3 FlexiForce Two force sensors are mounted on each of the foot of RO-PE-V to detect the ground contact of every step. The use of force sensors are more for slope walking which will be discussed with more details in Chapter 6. Fig. 3.5 shows the picture of the FlexiForce employed on RO-PE-V. Fig. 3.5 FlexiForce sensor employed on RO-PE-V. CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 18 FlexiForce from Tekscan was selected for it is small and light weight, such that they could be installed on the robot with minimum disturbance to the motion of the robot. FlexiForce is a resistive force sensor that changes resistance depending on the amount of force applied to the sensing area. When there is no load, the sensor has a high resistance of about 20M , while the resistance of the sensor would drop to range of K when it is loaded. To measure the contact force, through measuring the change in resistance, the sensor is connected in a potential divider as shown in Fig. 3.6 to output an analogue voltage signal for measurement. 5V 1k Output to analogue to digital converter FlexiForce Ground Fig. 3.6 Circuit for measuring the change in resistance in Flexiforce. 3.2 Actuators Actuators could be considered the most important component of a robot. They are the actual moving mechanisms that would allow a robot to perform an action, just like muscles on human. For a long time, servo motors from Japanese companies like Hitec, JR and Futaba have been dominating the market of actuators for small size robots, primarily because of their light weight and compactness in size. CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 19 However, as technology in robotics advances and with this research topic getting popular worldwide, competitors from other countries appear. Robotis from Korea is among the leading competitor, and Dynamixels DX-117 are employed as the only type of actuator on RO-PE-V. Fig. 3.7 gives a picture of DX-117 while Table 3.1 shows a comparison between HSR-5995 from Hitec and DX-117 from Robotis. Fig. 3.7 Dynamixel DX-117 from Robotis. Table 3.1 Comparisons between HSR-5995 from Hitec and DX-117 from Robotis. HSR-5995 (Hitec) DX-117 (Robotis) Max. Torque (Kg-cm) 30 37 Weight (g) 62 66 Speed (sec/60o) 0.12 0.129 Operating Angle (degree) 180 300 Supply Voltage (V) 4-6 12-16 CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS Link Through PWM generator RS485 Feedback No Yes Daisy Chain No Yes 20 The increase in torque and operating angle, the existence of feedback and daisy chain capability are the primary pull factors for the switch from Hitec motors to Robotis motors. The increase in torque and operating angle would allow RO-PE-V to have a higher payload and to execute more demanding actions. The daisy chain connections would minimize wires within the robot, cutting down weight and chances of wires being snipped by the mechanical structure in motion. And the availability of feedback in position and torque gives the possibility of implementing more sophisticated action algorithms, while the feedback in temperature and voltage could be used to protect the motors from overloading. DX-117 uses RS485 for communication with the controller of the robot, which is a standard protocol in the field of data acquisition, and it is this employed protocol that allows DX-117 to be daisy chained and a high transmission rate of up to 1Mbps. Each motor is given a unique ID in the setting phase, and because all of them are connected in the same lines, they will receive all instructions given by the controller. However, each instruction packet is led by the ID(s) of the desired receiving motor(s). Thus, the motors will only respond to instructions meant for them and ignoring the rest. CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 3.3 21 Computer Systems Computer systems serve as the brain of the robot. It makes decisions according to the environment information from the sensors’ feedback and a set of rules predetermined in the program. It disseminates its decisions in the form of instructions to the motors for execution. This sequence could be simple and does not require a very powerful processing unit. However, the processor on RO-PE-V would be tasked to perform image processing as well. This would be a demanding routine and the overall workload would require a powerful but compact processor. CRR3 CoolRoadRunnerIII from the PC104 family is selected for this application. It has a processing speed of 650MHz and is relatively compact in size. It is effectively a Pentium 3 computer in a small form factor. Real-time Windows is installed as the operating system for RO-PE-V with Microsoft Visual Studio as the programming environment. Table 3.2 lists some important specifications of CRR3. Table 3.2 Specifications of CRR3 from LIPPERT. Processor Speed 650MHz RAM 512MB Serial Communication 2 x RS232 USB 2 x USB1.1 compliant Supply Voltage 5V CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS Power Consumption [...]... positions of the centre of mass could be obtained, and through inverse kinematics, obtaining the joint trajectories Wasaeda University was the first to implement this control approach on a real robot [30] 2.2.3 Biologically inspired Passive dynamic walking is a form of walking behaviour that was discovered by Tad McGeer [6] It was shown that a passive walker could walk down a slope based on just gravity and. .. dominating the market of actuators for small size robots, primarily because of their light weight and compactness in size CHAPTER 3: SENSORS, ACTUATORS, AND COMPUTER SYSTEMS 19 However, as technology in robotics advances and with this research topic getting popular worldwide, competitors from other countries appear Robotis from Korea is among the leading competitor, and Dynamixels DX-117 are employed as... have physical meaning, for example, hip height and swing ankle height With these parameters, the computer on the robot can perform some computations to convert these information in Cartesian space to joint space, giving instructions that the actuators require The advantage of planning in Cartesian space is that path planning is on parameters with resultant of motions of many actuators, instead of going... linkages Fig 4.9 Designs of the thigh and shank links of RO-PE-V Fig 4.10 COSMOXpress analysis results on the thigh and shank links CHAPTER 4: MECHANICAL DESIGN 33 Ankle Joint There are two degrees of freedom required at the ankle joint, ankle pitch and ankle roll, and the design of this joint takes a similar form as the hip joint That is, the pitch and roll motors are connected in the same way This... stability because there is a larger area for the centre of gravity of the robot to fall within, thereby maintaining stability However, too large a foot would render the research uninteresting due to the lack of realism, and for the same reason, RoboCup has a foot size specification that the participating teams are supposed to adhere to Another foot design consideration is the need to accommodate force... aluminum alloy, which has a low density of about 2700Kg/m3, are used for the main skeleton of the robot Linkage designs are simple and connectors are positioned such that minimum dismantling is required in order to access and tighten any of them during use of the robot RoboCup has a set of robot specifications which participants has to adhere to when designing their robot Therefore, this poses as one of. .. version of the arm design Fig 4.3 Arm design using Perspex as linkages CHAPTER 4: MECHANICAL DESIGN 28 The use of Perspex indeed cuts down on weight, however, the tradeoff in using Perspex is a reduction in impact strength Due to the nature of the application, impact on the Perspex portion of the robot is unavoidable because of falling, and after one year of usage, a couple of the Perspex rods started... to break And therefore, the second version of arm design, which uses aluminum alloy, is used to replace the earlier design Fig 4.4 shows the new arm design Aluminum alloy Fig 4.4 New arm design using aluminum alloy as linkages 4.1.3 Body Design The main purposes of the body are to connect the limbs and to house the computer systems and the batteries A simple rectangular casing was designed for that,... physical movements upon receiving signals from the controller, and it is the combination of motions from many actuators that allows the humanoid robot to walk and (3) Computer Systems, which acts as the brain of the robot, take in signals from sensors, compute, come out with the necessary reaction plan and produce the signal to instruct the robot or, the actuators, the actions to be carried out CHAPTER... given to humanoid robots are rather simple and they usually operate in a structured area ASIMO from Honda [28] had demonstrated an encouraging level of intelligence by recognizing voice of people and moving around with people But still, there are much works to be done before humanoids could really be intelligent Chapter 3 Sensors, Actuators and Computer Systems A fully autonomous humanoid robot could ... hardware, (2) walking control and (3) artificial intelligence CHAPTER 2: LITERATURE REVIEW 2.1 Mechanical Design and Hardware In the area of mechanical design, one of the important areas is... normal gearbox remains the common selection by small size humanoid robots as they are usually integrated with motors as a compact package by the manufacturer and are much cheaper However, backlash... the actuators require The advantage of planning in Cartesian space is that path planning is on parameters with resultant of motions of many actuators, instead of going down to plan the paths of