1312 Part G Human-Centered and Life-Like Robotics 56.2.3 Different Sensors the robot will see a different world than the human.With Part G156 respect to behavior,placement of sensors on the head of Humanoid robots have made use of a variety of sen- the robot allows the robot to sense the world from a van- sors including cameras,laser range finders,microphone tage point that is similar to that of a human,which can arrays,lavalier microphones,and pressure sensors. w be important for finding objects that are sitting on a desk Some researchers choose to emulate human sensing by or table. selecting sensors with clear human analogs and mount- Prominent humanoid robots have added additional ing these sensors on the humanoid robot in a manner that sensors without human analogs.For example,Kismet mimics the placement of human sensory organs.As dis- used a camera mounted in its forehead to augment the cussed in Sect.56.6,this is perhaps most evident in the two cameras in its servoed eyes,which simplified com- use of cameras.Two to four cameras are often mounted mon tasks such as tracking faces.Similarly,versions of within the head of a humanoid robot with a configuration Asimo have used a camera mounted on its lower torso similar to human eyes. that looks down at the floor in order to simplify obstacle The justifications for this bias towards human-like detection and navigation during locomotion. sensing include the impact of sensing on natural human- robot interaction,the proven ability of the human senses 56.2.4 Other Dimensions of Variation to support human behavior,and aesthetics.For example, with respect to human-robot interaction,nonexperts can Other significant forms of variation include the size of sometimes interpret the functioning and implications of the robot.the method of actuation.the extent to which the a human-like sensor.such as a camera,more easily.Sim- robot attempts to appear like a human,and the activities ilarly,if a robot senses infrared or ultraviolet radiation, the robot performs. 56.3 Locomotion Bipedal walking is a key research topic in humanoid typically need to balance dynamically when walking robotics (see also Chap.16,Legged Robots,for a review bipedally. of this topic in the context of locomotion in general). Legged locomotion is a challenging area of robotics 56.3.1 Bipedal Locomotion research,and bipedal humanoid locomotion is espe- cially challenging.Some small humanoid robots are Currently the dominant methods for bipedal legged loco- able to achieve statically stable gaits by having large motion with humanoids make use of the zero-moment feet and a low center of mass,but large humanoids with point(ZMP)criterion to ensure that the robot does not a human-like weight distribution and body dimensions fall over [56.35].As discussed in detail in Chap.16,con- 1=0s t=2.5s 1=5s t=7.5s 1=17.5s t=15s t=12.5s t=10s Fig.56.9 HRP-2 walks on a slightly uneven surface
1312 Part G Human-Centered and Life-Like Robotics 56.2.3 Different Sensors Humanoid robots have made use of a variety of sensors including cameras, laser range finders, microphone arrays, lavalier microphones, and pressure sensors. Some researchers choose to emulate human sensing by selecting sensors with clear human analogs and mounting these sensors on the humanoid robot in a manner that mimics the placement of human sensory organs. As discussed in Sect. 56.6, this is perhaps most evident in the use of cameras. Two to four cameras are often mounted within the head of a humanoid robot with a configuration similar to human eyes. The justifications for this bias towards human-like sensing include the impact of sensing on natural human– robot interaction, the proven ability of the human senses to support human behavior, and aesthetics. For example, with respect to human–robot interaction, nonexperts can sometimes interpret the functioning and implications of a human-like sensor, such as a camera, more easily. Similarly, if a robot senses infrared or ultraviolet radiation, the robot will see a different world than the human. With respect to behavior, placement of sensors on the head of the robot allows the robot to sense the world from a vantage point that is similar to that of a human, which can be important for finding objects that are sitting on a desk or table. Prominent humanoid robots have added additional sensors without human analogs. For example, Kismet used a camera mounted in its forehead to augment the two cameras in its servoed eyes, which simplified common tasks such as tracking faces. Similarly, versions of Asimo have used a camera mounted on its lower torso that looks down at the floor in order to simplify obstacle detection and navigation during locomotion. 56.2.4 Other Dimensions of Variation Other significant forms of variation include the size of the robot, the method of actuation, the extent to which the robot attempts to appear like a human, and the activities the robot performs. 56.3 Locomotion Bipedal walking is a key research topic in humanoid robotics (see also Chap. 16, Legged Robots, for a review of this topic in the context of locomotion in general). Legged locomotion is a challenging area of robotics research, and bipedal humanoid locomotion is especially challenging. Some small humanoid robots are able to achieve statically stable gaits by having large feet and a low center of mass, but large humanoids with a human-like weight distribution and body dimensions t=17.5 s t=0 s t=15 s t=2.5 s t=12.5 s t=5 s t=10 s t=7.5 s Fig. 56.9 HRP-2 walks on a slightly uneven surface typically need to balance dynamically when walking bipedally. 56.3.1 Bipedal Locomotion Currently the dominant methods for bipedal legged locomotion with humanoids make use of the zero-moment point (ZMP) criterion to ensure that the robot does not fall over [56.35]. As discussed in detail in Chap. 16, conPart G 56.3
Humanoids 56.3 Locomotion 1313 Part G56. w Fig.56.12 The humanoid robot HRP-2P getting up from a lying- Fig.56.10 These robots from Delft,MIT and Cornell (left down position to right)are designed to exploit their natural dynamics when walking [56.36].(Image courtesy of Steven H.Collins) to this task.So far,a generic and rigorous new criterion has not been established. trol of the robot's body such that the ZMP sits within the As an example of an alternative mode of bipedal support polygon of the robot's foot ensures that the foot locomotion,some running robots have used controllers remains planted on the ground,assuming that friction is based on an inverted pendulum model to achieve sta- high enough to avoid slipping.The ZMP can be used to ble gaits.These methods change the landing positions plan motion patterns that make the robot dynamically to keep the robot dynamically stable [56.37].More re- stable while walking. cently,researchers have begun to use the principles Controllers based on the ZMP criterion try to fol-of bipedal passive-dynamic walkers to develop pow- low a planned sequence of contact states and are often ered bipedal walkers that walk with high efficiency unable to change the landing positions in real time inin a human-like way by exploiting natural dynam- response to lost contact.Current ZMP-based bipedal ics(Fig.56.10 [56.361). walking algorithms have difficulty handling unexpected perturbations,such as might be encountered with un- 56.3.2 Falling Down even natural terrain (Fig.56.9).Robots using ZMP differ from human locomotion in significant ways.For exam- A human-scale robot should expect to fall from time ple,unlike people,robots using ZMP typically do not to time in realistic conditions.A humanoid robot may exploit the natural dynamics of their legs,or control the fall down due to a large disturbance even if the motion impedance of their joints. is planned carefully and a sophisticated feedback con- Augmenting ZMP-based control is currently an ac- troller is applied to the robot.In this event,the robot tive area of research.As will be discussed in Sect.56.5 could be damaged significantly during a fall,and could on whole-body activities researchers are working to also damage the environment or injure people who are integrate manipulation and bipedal locomotion.For ex-nearby.An important area of research is how to control ample,when a robot walks while grasping a handrail,the robot's fall in order to gracefully recover or minimize the contact could potentially increase the stability of the damage.The Sony QRIO can control its falling motions robot,but the ZMP criterion does not easily generalize in order to reduce the impact of touch down [56.38],al- Fig.56.11 Example of controlled falling-down motion
Humanoids 56.3 Locomotion 1313 Fig. 56.10 These robots from Delft, MIT and Cornell (left to right) are designed to exploit their natural dynamics when walking [56.36]. (Image courtesy of Steven H. Collins) trol of the robot’s body such that the ZMP sits within the support polygon of the robot’s foot ensures that the foot remains planted on the ground, assuming that friction is high enough to avoid slipping. The ZMP can be used to plan motion patterns that make the robot dynamically stable while walking. Controllers based on the ZMP criterion try to follow a planned sequence of contact states and are often unable to change the landing positions in real time in response to lost contact. Current ZMP-based bipedal walking algorithms have difficulty handling unexpected perturbations, such as might be encountered with uneven natural terrain (Fig. 56.9). Robots using ZMP differ from human locomotion in significant ways. For example, unlike people, robots using ZMP typically do not exploit the natural dynamics of their legs, or control the impedance of their joints. Augmenting ZMP-based control is currently an active area of research. As will be discussed in Sect. 56.5 on whole-body activities researchers are working to integrate manipulation and bipedal locomotion. For example, when a robot walks while grasping a handrail, the contact could potentially increase the stability of the robot, but the ZMP criterion does not easily generalize Fig. 56.11 Example of controlled falling-down motion Fig. 56.12 The humanoid robot HRP-2P getting up from a lyingdown position to this task. So far, a generic and rigorous new criterion has not been established. As an example of an alternative mode of bipedal locomotion, some running robots have used controllers based on an inverted pendulum model to achieve stable gaits. These methods change the landing positions to keep the robot dynamically stable [56.37]. More recently, researchers have begun to use the principles of bipedal passive-dynamic walkers to develop powered bipedal walkers that walk with high efficiency in a human-like way by exploiting natural dynamics (Fig. 56.10 [56.36]). 56.3.2 Falling Down A human-scale robot should expect to fall from time to time in realistic conditions. A humanoid robot may fall down due to a large disturbance even if the motion is planned carefully and a sophisticated feedback controller is applied to the robot. In this event, the robot could be damaged significantly during a fall, and could also damage the environment or injure people who are nearby. An important area of research is how to control the robot’s fall in order to gracefully recover or minimize damage. The Sony QRIO can control its falling motions in order to reduce the impact of touch down [56.38], alPart G 56.3
1314 Part G Human-Centered and Life-Like Robotics there is also the issue of getting back up again [56.40] Part G56. (Fig.56.12). 56.3.3 Sensing for Balance Rubber bushing w Bipedal walking needs to be robust to unexpected dis- turbances encountered during the execution of planned walking patterns.In these situations,walking can sometimes be stabilized with feedback control and 6-axis force sensor appropriate sensing.Many humanoid robots,such as Rubber sole Honda's Asimo,make use of accelerometers,gyro- Fig.56.13 Example of a humanoid foot structure for legged loco- scopes,and six-axis force/torque sensors to provide motion that uses compliance and force/torque sensing feedback to the robot during locomotion. Force/torque sensors have long been applied to ma- nipulators for the implementation of force control,but force/torque sensors with sufficient robustness to handle foot impact for a full-size humanoid robot are relatively new.When the foot of the robot touches down,the foot receives an impact which can disturb its walking.This impact can be rather large,especially when the robot is walking quickly.Some feet now incorporate a spring and damper mechanism as shown in Fig.56.13 in order to mitigate these problems.As with many other aspects of bipedal humanoid locomotion,foot design is currently an open problem. 56.3.4 Localization and Obstacle Detection In order for a humanoid robot to walk in unmod- Fig.56.14 Asimo and artificial landmarks on the floor eled environments.localization and obstacle detection are essential.Wheeled robots encounter similar is- though it is of a relatively small size (which simplifies sues while navigating,but full bipedal humanoids have the problem).Fujiwara et al.developed a falling mo- more-specialized requirements.For example,bipedal tion controller for a human-size humanoid robot that is humanoids have the ability to control contact with the falling backwards [56.39].Figure 56.11 shows an exam- world through their highly articulate legs. ple of a controlled falling motion.The general problem Artificial landmarks can simplify localization.As is still very much an active area of research.Similarly, shown in Fig.56.14,Honda's Asimo uses a camera mounted on its lower torso that looks down at the floor to find artificial markers for position correction [56.41]. Accurate positioning is important for long-distance nav- igation and stair climbing,since slippage usually occurs while walking and accumulated positional and direc- tional errors can lead to severe failures. Obstacle avoidance is also an important function for locomotion.Disparity images generated by stereo vision have been utilized for this purpose.For example,the plane segment finder [56.42]developed by Okada et al. helps detect traversable areas.Figure 56.15 shows the re- sult of detecting clear areas of the floor plane appropriate for gait generation. Humanoids require a great deal of computation Fig.56.15 Plane segment finder for detecting traversable floor area due to the need for sophisticated sensing and con-
1314 Part G Human-Centered and Life-Like Robotics 6-axis force sensor Rubber sole Rubber bushing Fig. 56.13 Example of a humanoid foot structure for legged locomotion that uses compliance and force/torque sensing Fig. 56.14 Asimo and artificial landmarks on the floor though it is of a relatively small size (which simplifies the problem). Fujiwara et al. developed a falling motion controller for a human-size humanoid robot that is falling backwards [56.39]. Figure 56.11 shows an example of a controlled falling motion. The general problem is still very much an active area of research. Similarly, Fig. 56.15 Plane segment finder for detecting traversable floor area there is also the issue of getting back up again [56.40] (Fig. 56.12). 56.3.3 Sensing for Balance Bipedal walking needs to be robust to unexpected disturbances encountered during the execution of planned walking patterns. In these situations, walking can sometimes be stabilized with feedback control and appropriate sensing. Many humanoid robots, such as Honda’s Asimo, make use of accelerometers, gyroscopes, and six-axis force/torque sensors to provide feedback to the robot during locomotion. Force/torque sensors have long been applied to manipulators for the implementation of force control, but force/torque sensors with sufficient robustness to handle foot impact for a full-size humanoid robot are relatively new. When the foot of the robot touches down, the foot receives an impact which can disturb its walking. This impact can be rather large, especially when the robot is walking quickly. Some feet now incorporate a spring and damper mechanism as shown in Fig. 56.13 in order to mitigate these problems. As with many other aspects of bipedal humanoid locomotion, foot design is currently an open problem. 56.3.4 Localization and Obstacle Detection In order for a humanoid robot to walk in unmodeled environments, localization and obstacle detection are essential. Wheeled robots encounter similar issues while navigating, but full bipedal humanoids have more-specialized requirements. For example, bipedal humanoids have the ability to control contact with the world through their highly articulate legs. Artificial landmarks can simplify localization. As shown in Fig. 56.14, Honda’s Asimo uses a camera mounted on its lower torso that looks down at the floor to find artificial markers for position correction [56.41]. Accurate positioning is important for long-distance navigation and stair climbing, since slippage usually occurs while walking and accumulated positional and directional errors can lead to severe failures. Obstacle avoidance is also an important function for locomotion. Disparity images generated by stereo vision have been utilized for this purpose. For example, the plane segment finder [56.42] developed by Okada et al. helps detect traversable areas. Figure 56.15 shows the result of detecting clear areas of the floor plane appropriate for gait generation. Humanoids require a great deal of computation due to the need for sophisticated sensing and conPart G 56.3
Humanoids 56.4 Manipulation 1315 trol.Customized computational hardware may help real time from the stereo cameras.This real-time mitigate this problem.For example,Sony's humanoid vision system has been used to detect floor ar- Part robot QRIO is equipped with a field-programmable eas,stair steps,and obstacles for navigation [56.43, 0 gate array (FPGA)to generate disparity maps in 44]. 3 56.4 Manipulation Hands and arms are the main interfaces with which one at the elbow,and three at the wrist.The use of seven humans act on the world around them.Manipulation DOFs results in a redundant degree of freedom with research within humanoid robotics typically focuses on respect to the six-DOF pose of the hand.To reduce me- the use of anthropomorphic arms,hands,and sensors chanical complexity,humanoid robot arms sometimes to perform tasks that are commonly performed by peo-have fewer than seven DOFs,for example,ARMAR- ple.Several chapters of the handbook relate to these III and Justin have seven-DOF arms,Cog and Domo goals,including Chap.24(Visual Servoing and Visual have six-DOF arms,and Asimo has five-DOF arms Tracking),Chap.26 (Motion for Manipulation Tasks),(Fig.56.16)[56.45,46]. and Chap.28(Grasping). Humanoid robot hands tend to vary more in their design (see Chap.15,Robot Hands).The human hand is 56.4.1 The Arm and Hand highly complex with over 20 DOFs(i.e.,approximately four DOFs per finger and a five-DOF thumb)in a very The kinematics of humanoid robot arms emulate the compact space with a compliant exterior,dense tactile human arm,which can be approximated by seven de-sensing,and muscular control.If a robot hand is to be grees of freedom(DOFs),with three at the shoulder, mounted on a robot arm,there are additional constraints in terms of the mass of the robot hand,since the hand sits at the end of the arm and must be efficiently moved in Fig.56.16 The humanoid robot Justin has two seven-DOF torque-controlled arms (DLR-Lightweight-Robot-III),and two 12-DOF hands (DLR-Hand-II).Justin's body is larger Fig.56.17 3-D object recognition by HRP-2 using versatile than a human's volumetric vision
Humanoids 56.4 Manipulation 1315 trol. Customized computational hardware may help mitigate this problem. For example, Sony’s humanoid robot QRIO is equipped with a field-programmable gate array (FPGA) to generate disparity maps in real time from the stereo cameras. This real-time vision system has been used to detect floor areas, stair steps, and obstacles for navigation [56.43, 44]. 56.4 Manipulation Hands and arms are the main interfaces with which humans act on the world around them. Manipulation research within humanoid robotics typically focuses on the use of anthropomorphic arms, hands, and sensors to perform tasks that are commonly performed by people. Several chapters of the handbook relate to these goals, including Chap. 24 (Visual Servoing and Visual Tracking), Chap. 26 (Motion for Manipulation Tasks), and Chap. 28 (Grasping). 56.4.1 The Arm and Hand The kinematics of humanoid robot arms emulate the human arm, which can be approximated by seven degrees of freedom (DOFs), with three at the shoulder, Fig. 56.16 The humanoid robot Justin has two seven-DOF torque-controlled arms (DLR-Lightweight-Robot-III), and two 12-DOF hands (DLR-Hand-II). Justin’s body is larger than a human’s one at the elbow, and three at the wrist. The use of seven DOFs results in a redundant degree of freedom with respect to the six-DOF pose of the hand. To reduce mechanical complexity, humanoid robot arms sometimes have fewer than seven DOFs, for example, ARMARIII and Justin have seven-DOF arms, Cog and Domo have six-DOF arms, and Asimo has five-DOF arms (Fig. 56.16) [56.45, 46]. Humanoid robot hands tend to vary more in their design (see Chap. 15, Robot Hands). The human hand is highly complex with over 20 DOFs (i. e., approximately four DOFs per finger and a five-DOF thumb) in a very compact space with a compliant exterior, dense tactile sensing, and muscular control. If a robot hand is to be mounted on a robot arm, there are additional constraints in terms of the mass of the robot hand, since the hand sits at the end of the arm and must be efficiently moved in Fig. 56.17 3-D object recognition by HRP-2 using versatile volumetric vision Part G 56.4