In the literature widespread use of the term “human-like” with respect to locomotion or gait can be found. It is often used to assess a presented, specific locomotion performance implying that human motion is perceived as optimal or, at least, worthy of imitation without providing any evidence for such a statement, as for instance in [108]. Human movements in general, however, are not by default optimal, as diverse counter-examples in biology prove [124].
Let us try to shed more light on the exemplary uses of this term [137], noting before that unfortunately the overwhelming majority of authors take a general but not rigorously specified understanding of “human-like locomotion” for granted and do not attempt to explain the use of this term in their work (e.g. [163, 147]). However, the different works may have very individual interpretations and definitions of this term. Others implicitly provide hints for their individual conception of the term. McGeer presents some snapshots of the robot motion sequence in [97] and formulates: “The gait of figure 2 is obviously anthropomorphic” (p. 1644). Additionally, he compares his and the robot model’s step frequency to support his assessment. Visual comparison of snapshots is a popular tool also in Schultz and Mombaur [148]. The obtained numerical results are compared with snapshots of a professional athlete’s running gait. “The periodic running motion looks very natural, (compare the corresponding animation at our website)” ([148], p. 789). Further it is argued based on the visual comparison, that “[...], the qualitative match is very good” ([148], p. 790). Joint angle histories are not compared; instead biomechanical gait characteristics such as duty factor, step length, contact/friction forces and the vertical center of gravity motion patterns are computed and compared to those known from a human. In [71], Iida et al. compare the sagittal joint angle histories obtained from the simulation and real robot model to human joint kinematics. Partial similarities in only two of the three sagittal leg joints, good agreement with the vertical CoM patterns and low similarities of the GRF are considered as sufficient to term the obtained bipedal locomotion “human-like”. Besides, it is pointed out that self-stabilization of a gait is seen as a further important property of human locomotion. According to Ogura et al. walking with a constant waist height and bent knees is not very “human-like”: “The ability to walk with stretched knees is an important quality that a humanoid robot should possess in order for it to mimic human motion” ([119], p. 3976). The heel-contact and toe-off are considered as other important characteristics of human walking. “Thus, if the robot realizes not only walking with stretched knees but also heel-contact and toe-off motions, it can be said that its walking style, in comparison to those of other humanoid robots, is more similar to that of humans” ([119], p. 3976). As proof of concept it is shown that the GRF of the robot agree very well in pattern and peaks with those collected from a human subject. Collins and Ruina [24] consider energy consumption, computed by the well known cost of transport (CoT) (cf. Section 4.5), as a measure to rate the degree of “human-likeness”. A robot walking at low CoT, like the Cornell biped, is regarded as quite “human-like”. For others, as in [114], a characteristic of “natural human-like motion” in general is the exploitation of the passive body dynamics.
This widely varying use of the term “human-like” suggests that a general understanding and common definition for the term “human-like locomotion” is missing. This results in different views regarding locomotion performance evaluation, without a generally accepted understanding of the neuromechanics and core functionality underlying human locomotion. The problem of the lack of a taxonomy leads also to the issue of the lack of benchmarks. For the purpose of
benchmarks and valid comparability of results among the different research groups, it should be therefore aimed at developing a comprehensive common definition for the term “human- like locomotion” that will provide valuable guidelines, significantly enhance the progress in the field of humanoid robot locomotion and also enable applications to human health. Solving this problem requires combined efforts of biomechanists and roboticists and probably at least one single PhD thesis dedicated only to this topic. Therefore, in order to increase the understanding of the studies performed within the scope of this thesis and of the related use of this term, we will briefly discuss closely interrelated questions, namely:
• Which are the main features of human motion performance that are desirable to be realized by robots [134]?
• Further, how are these features embodied and related to each other [134]?
Human motion capture data offers a wealth of data ranging from GRF to Electromyogra- phy (EMG) and kinematic data. But the essential question is which of all these reference data actually capture the necessary information about the human motion dynamics. So far, all these data seem to contain redundant information. For instance, the GRF not only provide information about the patterns and forces. Using the GRF we can compute also the duty factor of the motion and in this way get an idea of the gait type. Additionally, the GRF let us derive the course of the CoM and consequently reveal the altitude difference of the CoM. EMG data, on the other hand, capture the electrical activity of skeletal muscles. There is controversy in the biome- chanics community whether it is also possible to estimate from EMG data the forces produced by the muscles. Recently, Sartori et al. reported on the feasibility of such estimation [145]. They demonstrated that by considering all the DoFs of the joints spanned by a muscle it is possible to estimate the muscle forces. Previous research, that had stated the opposite, used models about a single DoF. Despite this interesting research result, so far EMG data are not considered as reliable source of information to be used for realizing similar motions on robots. Another very popular opportunity, often applied in robotics and graphics research, is the use of kinematic data.
The question remains which of these above sources may play a role to support the development of a motion controller that produces the qualitatively best possible motions on a robot. For example, the relation of CoM or ground contact dynamics to joint angle trajectories is not unique, i. e., the same joint angle configurations can lead to different CoM locations or GRF. But there must be some dependency that is worth being analyzed. Further, it is not clear how joint angle courses affect the energy consumption.
While for the moment it does not appear to be possible to formulate a generally accepted answer to the above posed questions, it seems that we still have a clear idea of what we would consider as comparable to human motion. This can be strongly observed during new demon- strations of ASIMO’s locomotion capabilities [2], which generate a vivid debate about their “human-likeness”. Apparently, both kinematic constraints and high dynamic mobility, expressed by clear ground clearances and long flight phases, are important features for our judgment on the degree of resemblance to human jogging motion. Low energy consumption is a further criterion for considering motions performed by robots as comparable to human capabilities. Due to the open issues already discussed and the missing knowledge of the truly underlying mechanisms of human locomotion it can be agreed upon that we are far away from claiming a motion being human-like. Instead, the goal should be to demonstrate improved locomotion performance and to present novel insights by studies in different related areas. In this thesis the
term “human-like” with respect to locomotion performance is avoided as much as possible and, if used, shall indicate any prevalent similarities to the human sagittal leg joint angle courses and ground contact dynamics provided a low energy-consumption. The simulated and real robot dynamics will be studied in detail with respect to the aforementioned important characteristics known from human gait analyses (cf. Chapter 4 and 5).