Robotics-based Synthesis of Human Motion - CiteSeerX


Robotics-based Synthesis of Human Motion Khatib, O.a , Demircan, E.a , De Sapio, V.a,b , Sentis, L.a , Besier, T.c , Delp, S.d a

Artificial Intelligence Laboratory, Stanford University, Stanford, CA 94305, USA {khatib, emeld, lsentis} b

Sandia National Laboratories, Livermore, CA 94551, USA [email protected]


Human Performance Laboratory, Stanford, CA 94305, USA [email protected]


Neuromuscular Biomechanics Laboratory, Stanford, CA 94305, USA [email protected]

Abstract The synthesis of human motion is a complex procedure that involves accurate reconstruction of movement sequences, modeling of musculoskeletal kinematics, dynamics and actuation, and characterization of reliable performance criteria. Many of these processes have much in common with the problems found in robotics research. Task-based methods used in robotics may be leveraged to provide novel musculoskeletal modeling methods and physiologically accurate performance predictions. In this paper, we present (i) a new method for the real-time reconstruction of human motion trajectories using direct marker tracking, (ii) a task-driven muscular effort minimization criterion and (iii) new human performance metrics for dynamic characterization of athletic skills. Dynamic motion reconstruction is achieved through the control of a simulated human model to follow the captured marker trajectories in real-time. The operational space control and real-time simulation provide human dynamics at any configuration of the performance. A new criteria of muscular effort minimization has been introduced to analyze human static postures. Extensive motion capture experiments were conducted to validate the new minimization criterion. Finally, new human performance metrics were introduced to study in details an athletic skill. These metrics include the effort expenditure and the feasible set of operational space accelerations during the performance of the skill. The dynamic characterization takes into account skeletal

kinematics as well as muscle routing kinematics and force generating capacities. The developments draw upon an advanced musculoskeletal modeling platform and a task-oriented framework for the effective integration of biomechanics and robotics methods. Index Terms task-space framework, human motion analysis, robotics, musculoskeletal dynamics, human animation, operational space formulation

1. I NTRODUCTION In the field of robotics, the motivation to emulate human movement is driven by the proliferation of humanoid robots and the desire to endow them with human-like movement characteristics (Nakamura et al.,2003). Inspired by human behaviors, our early work in robot control encoded tasks and diverse constraints into artificial potential fields capturing human-like goal-driven behaviors (Khatib and Le Maitre,1978). This concept was later formalized in the task oriented operational space dynamic framework (Khatib,1986; Khatib,1987). More recently, this formulation was extended to address whole-body control of humanoid robots and successfully validated on physical robots (Khatib et al.,2004). The framework provides multi-task prioritized control architecture allowing the simultaneous execution of multiple objectives in a hierarchical manner, analogous to natural human motion (see Fig. 1). One of the major difficulties associated with the prediction and synthesis of human movement is redundancy resolution. Whether the goal is to gain an understanding of human motion or to enable synthesis of natural motion in humanoid robots a particularly relevant class of movements involves targeted reaching. Given a specific target the prediction of kinematically redundant limb motion is a problem of choosing one of a multitude of control solutions all of which yield kinematically feasible solutions. It has been observed that humans resolve this redundancy problem in a relatively consistent manner (Kang et al.,2005; Lacquaniti and Soechting,1982). For this reason, mathematical models have proven to be valuable tools for motor control prediction (Hermens and Gielen,2004; Vetter et al.,2002). These models frequently characterize some 2

Fig. 1.

Task-space framework allows the articulated skeleton to maintain balance while accomplishing a manual task

(Khatib et al.,2004).

element of musculoskeletal effort. Robotics-based effort models frequently utilize quantities that are derivable purely from skeletal kinematics and that are not specific to muscle actuation. It is thus useful to consider an analogous measure that encodes information about the overall musculoskeletal system to account for muscle actuation and its redundancy. Activation, which represents the normalized exertion of muscles, provides a natural starting point for constructing such a measure. Specifically, the magnitude of muscle activation vector has been used as an optimization criterion in both static and dynamic optimizations (Thelen et al.,2003). The utilization of a model-based characterization of muscle systems, which accounts for muscle kinematic and strength properties, is critical to authentically simulating human motion since human motions are frequently linked by physiological constraints. In this paper, a robotic approach for the synthesis of human motion using a task-space framework is presented. For this purpose, the direct marker control for human motion reconstruction, a criterion of task-driven effort minimization and new metrics for dynamic characterization of human performance were introduced. The result is a dynamic biomechanical profile of human performance that facilitates the modeling of human motion. These approaches were tested through extensive motion capture experiments on human subjects including a martial


art master and a professional football player. The results showed that these skillful practitioners tend to minimize the muscular effort while following the lines of maximum feasible accelerations when performing a task. These results support the prediction that task-driven human motions emerge from the use of biomechanical advantage of the human musculoskeletal system under physiological constraints. 1.1 Task Dynamic Behavior and Control For a given desired whole-body task of a human-like robot, the motion behaviors should be specified to be controlled during the execution of the motion. Hand location, balance, effort minimization, and obstacle and joint limit avoidance are common choices, but the exhaustive list depends upon the motion to be performed. Considering each behavior as an independent task, the number of degrees of freedom describing each task is typically less than the number of joints in the robot. For these situations, there are multiple ways of performing the task. This redundancy is labeled in solutions as the posture space of the task, containing all possible motions that do not affect task performance (Khatib et al.,2004). As such, other tasks may be controlled by selectively choosing the path within the posture space. In this section, the dynamic model of the task/posture decomposition and the model describing the motion of the subtask within the posture space (Khatib et al.,2004) are reviewed. Combination of these two models provides a control structure that compensates for the dynamics in both spaces, significantly improving performance and responsiveness for multiple tasks. A task can be defined to be any formal description of desired activity that can be explicitly represented as a function of the joint coordinates, q, q˙ and q¨. Multiple tasks, xi ’s, can be combined into a single task definition in a higher dimensional space, as long as they are kinematically consistent with each other. The task coordinates are denoted by xt = xt (q). The joint space equations of motion can be expressed as, A(q)¨ q + b(q, q) ˙ + g(q) = Γ,


where q is the n × 1 vector of generalized coordinates, A(q) is the n × n mass matrix, b(q, q) ˙ is 4

the n × 1 vector of centrifugal and Coriolis terms, g(q) is the n × 1 vector of gravity terms, and Γ is the n × 1 vector of generalized control forces (torques). For conciseness we will often refrain from explicitly denoting the functional dependence of these quantities on q and q. ˙ The Jacobian matrix associated with the task, xt , is denoted by Jt (q). The task dynamic behavior can be obtained by projecting the skeletal dynamics (1) into the space associated with the task, using the generalized inverse of the Jacobian, J¯t . This generalized inverse of the Jacobian has been showed to be unique and dynamically consistent (Khatib,1987; Khatib,1995) and given by, J¯t , A−1 JtT (Jt A−1 JtT )−1 .


The dynamic behavior associated with the task, xt can be obtained by, J¯tT (A¨ q + b + g = Γ) ⇒ Λt x¨t + µt + pt = Ft .


In this space, Λt is the m × m task inertia matrix, and µt , pt , and Ft are respectively the centrifugal and Coriolis forces, gravity effect, and generalized force acting along the direction of the task, xt . This process provides a description of the dynamics in task coordinates rather than joint space coordinates (while joint space coordinates are still present in (3), the inertial term involves task space accelerations rather than joint space accelerations). The control framework defined in terms of the relevant task coordinates, xt , can be represented using a relevant operational space force, Ft , acting along the same direction. The forces acting along given task coordinates can be mapped to a joint torque, Γtask , by the relationship, Γtask = JtT Ft .


For a given task, there is a unit inertial behavior, It x¨t = Ft∗ and Ip x¨p = Fp∗ . The nonlinear dynamic control force of the task, Ft , is given by, b t Ft∗ + µ bt + pbt , Ft = Λ 5


where b. denotes the estimates of the components of the dynamic models and F ∗ is the desired force. The generalized torque/force relationship (Khatib,1987; Khatib,1995) allows for the decomposition of the total torque into two dynamically decoupled torque vectors: the torque corresponding to the commanded task behavior and the torque that only affects posture behaviors in the null space provided by the kinematic redundancy of the musculoskeletal system, Γ = Γtask + Γposture .


The operational space formulation determines the torque component for the task to compensate for the dynamics in the task space, Γ = Γtask + Γposture = JtT Ft + NtT Γp ,


where Nt is the null space associated with the task. Dynamically consistent posture control guarantees posture behaviors to be performed without projecting any acceleration onto the task (Khatib,1995). Any acceleration associated with the posture that would affect the task is filtered by its null space, Nt . Postures can be represented by minimal sets of independent posture coordinates, Γp = JpT Fp .


The task consistent posture Jacobian Jp|t can be defined through the relation, Jp|t = Jp Nt .


The task description and whole-body dynamic control through prioritization can be obtained by, T [A¨ q + b + g = Γtask + Γposture ] ⇒ Λp|t x¨p|t + µp|t + pp|t = Fp|t , J¯p|t


and the force/torque relationship, T Γposture = Jp|t Fp|t .



Using these dynamic behavior models, a dynamically decoupled control to perform both tasks can be formulated. The control force for the decoupled system, Fp|t , is given by, b p|t F ∗ + µ Fp|t = Λ bp|t + pbp|t , p|t


∗ where Fp|t is the desired force for the decoupled system.

Using the task-dependent torque decomposition and the force/torque relationship, the resulting control torque, Γ, is, T Γ = Γtask + Γposture = JtT Ft + Jp|t Fp|t .


The task can be controlled by a task field Ut that determines the desired behavior by its gradient Ft∗ = −∇xt Ut . Similarly, the posture behavior Fp∗ can be specified by a posture field Ut . In the study of human motion, the strategies humans follow to perform skills can be expressed by these energy potentials. 1.2 Human Motion Reconstruction by Direct Marker Control The motion capture is an effective tool to investigate human kinematics in a given motion. However, a number of post processing steps need to be performed to convert the raw marker positions into useful kinematic data. The most significant step is to convert the marker trajectories, x, x˙ and x¨, into joint space trajectories, q, q˙ and q¨. This has commonly been done using inverse kinematic techniques. As an alternative to performing inverse kinematics on marker data, a new algorithm is proposed to reconstruct human movement through direct control of optical marker trajectory data (Demircan et al.,2008). This approach which dynamically tracks the markers using the task-level control (Khatib and Burdick,1987) and the prioritized control (Khatib et al.,2004; Sentis and Khatib,2005) frameworks is referred to as direct marker control. The direct marker control is achieved by mapping a scaled dynamic human model to the experimental marker locations in Cartesian space and simulating it along the desired trajectories in real-time. In order to accurately reconstruct human motion, the direct marker control algorithm solves the problem of redundancy (i.e. much marker position data than needed to resolve the joint 7

angles) and ensures marker decoupling (markers on the same body link are rigidly constrained to each other and the relative motion between markers on adjacent links is limited by the freedom in the connecting joints) by grouping the markers into independent subsets and forming a hierarchy of tasks associated with each subset. In this marker space, a priority is assigned to each subset task and the tasks that have lower priorities in the hierarchy are projected into the null space of the tasks that have higher priorities. This process is recursively iterated and the human model is tracked to the desired motion configurations. In order to have kinematically consistent motion patterns, the human model is scaled to the subject’s antropometry. Kinematically correct human model is then simulated in real-time to generate the motion dynamics at any state of the performance. The direct marker control algorithm constitutes an effective tool by extending our dynamic environment to identify the characteristics describing natural human motion which can be mapped into humanoid robots for real-time control and analysis. Direct Marker Control Framework: For the purpose of using motion capture systems to reconstruct a human motion, the task/posture decomposition used in the operational space method constitutes a natural decomposition for dealing with marker data, thus avoiding the performance of inverse kinematics. This decomposition allows us to represent the dynamics of a simulated human subject in a relevant task space that is complemented by a posture space (7). For an arbitrary number of tasks, the torque decomposition (13) can be generalized to, Γ = JtT1 Ft1 + JtT2 |t1 Ft2 |t1 + · · · + JtTn |tn−1 |···|t1 Ftn |tn−1 |···|t1 .


In the direct marker control application, task space is defined as the space of Cartesian coordinates for the motion capture markers. However, marker trajectories obtained through motion capture are not independent. To accommodate for the motion dependencies, the markers are grouped into independent subsets, m1 , ..., mn , where mi denotes the task for a particular marker subset. At the end of the recursive process of building a marker space defined by a


Fig. 2.

Scaled human model of a tai chi master. Markers of the right shoulder and the left wrist are selected to form the first

marker set to be controlled (dark spheres). The second subset is formed by the left elbow and the right wrist markers (light spheres) (Demircan et al.,2008). The musculoskeletal model was derived from models of the upper extremity (Holzbaur et al.,2005) and lower extremity (Delp et al.,1990).

hierarchy of decoupled marker tasks, we obtain, T T T Γ = Jm F + Jm F + · · · + Jm F . n |mn−1 |···|m1 mn |mn−1 |···|m1 1 m1 2 |m1 m2 |m1


The Jacobian and the force associated with marker space are deduced from the above equation as follows,

      J⊗ ,     

Jm1 Jm2 |m1 .. .

         

     F⊗ ,     


Jmn |mn−1 |···|m1

 Fm1 Fm2 |m1 .. .

         


Fmn |mn−1 |···|m1

The overall control torque defined in marker space is then, Γ = J⊗T F⊗ .


Experimental Validation: To test the direct marker control algorithm, a series of movements performed by a tai chi master were captured using an 8-camera motion capture system. The motion was then reconstructed in the control and simulation framework, SAI (Khatib et al.,2002)


Fig. 3.

Tracked and goal trajectories of markers. The tracked trajectories (solid lines) are shown for markers attached to the

wrist, shoulder, and elbow segments. The tracked trajectories closely match the goal (experimental) trajectories (dotted crosses) (Demircan et al.,2008).

by tracking the marker trajectories in real-time. Prior to tracking, our existing human model which consists of 25 joints, was first scaled to match the anthropometry of the tai chi master. The human motion reconstruction was then executed using subsets of decoupled marker trajectories. Fig. 2 illustrates the scaled musculoskeletal model together with the marker subsets selected for direct control. The commanded and tracked positions of the controlled markers (Fig. 3), as well as the joint angles (Fig. 4), were recorded during real-time simulation. The results demonstrated the effectiveness of the direct marker control algorithm in ensuring smooth tracking of marker trajectories and for the extraction of joint angles without inverse kinematics computations. An analysis on the bounds of the joint space errors can be performed using the Jacobian 10

Fig. 4.

Right arm joint angles obtained through direct control of marker data. Smooth joint space trajectories are obtained as a

natural output of the marker tracking methodology (Demircan et al.,2008).

Fig. 5. Margin of marker position errors and margin of joint angle errors over the trajectory. Joint angle error magnitudes show a stable variation over the trajectory, thus ensuring well bounded errors on the joint angles (Demircan et al.,2008).

associated with the marker space, by: ∆x⊗ = J⊗ ∆q.


Figure 5 shows the margin of marker position errors and the margin of joint angle errors respectively. Maximum and minimum joint angle error magnitudes vary stably over the trajectory, suggesting well bounded errors on the joint angles.


2. P HYSIOLOGY- BASED P OSTURE F IELD AND U NIFIED M USCULOSKELETAL TASK /P OSTURE C ONTROL In learned tasks, humans exploit the biomechanical advantage of their musculoskeletal system by positioning their body in a way to reach the best configuration while continuously adjusting their posture. For a given task, our basic prediction is that the strategy for the transmission of this biomechanical advantage in the presence of physiological capacities corresponds to a minimization of the energy associated with muscular effort. Consider a limb with one muscle and let m designate the force generated by this muscle. When adding the effects of other muscles, a weighting coefficient, c is used to account for the force generating capacity of a given muscle. The energy, E, associated with the effort produced by this single muscle as a function of f (c, m) can be defined by, E = cm2


For a multi-muscle musculoskeletal system, the muscular forces take the form of a vector. The corresponding joint torque associated with m is given by the relationship, Γ = LT m,


where L is the r × n muscle Jacobian matrix (moment arms) for a system of n joints and r muscles. As a criteria for natural human motion, the human posture is continuously adjusted to reduce muscular effort (Khatib et al.,2004). For a given task, the muscle effort measure, E(q), can be given by the constituent terms using the generalized operational space force, F , and the relationship (4), as, E(q) = F T J(q)(LT Nc2 L)−1 J T (q)F,


where Nc is the r × r muscle capacity matrix and relates to the weighting coefficient, c, by the relationship, c = Nc−2 . 12


In (21) the terms inside the parentheses represent a measure of the net capacity of the muscles. This is a combination of the force generating kinetics of the muscles as well as the mechanical advantage of the muscles, as determined by the muscle moment arms. In dynamic skills, inertial forces are part of the effort and are taken into account accordingly. Equation 21 represents a generalization of the joint decoupled measure used in (Khatib et al.,2004), which projected muscle strength capacities to the joint level in a decoupled manner. Consequently, the cross-joint coupling associated with multi-articular muscles was ignored. The energetic cost measure in (21) properly accounts for multi-articular muscle coupling in the musculoskeletal system. For posture-based analysis the static form of the instantaneous muscle effort measure can be constructed by noting that overall torque is reduced to g, in the absence of external loads and (21) takes the static form as, E(q) = g T (q)(LT Nc2 L)−1 g(q).


Human Model: In order to evaluate the posture-based muscle effort criterion, a musculoskeletal model must be implemented. Fidelity in predicting muscle lines of action and moment arms was an important requirement for this model. In particular, proper kinematics of the shoulder complex are critical in generating realistic muscle paths and associated joint moments of the upper limb. For this reason the upper extremity model of (Holzbaur et al.,2005) has been employed, with some modification, in this work. This model is characterized by coupled motion between the shoulder girdle and the glenohumeral joint. An extensive analysis of this model, in particular the impact of shoulder girdle motion on the muscle routing kinematics and moment arms about the glenohumeral joint, is provided in (De Sapio et al.,2006). The model, consisting of a constrained shoulder complex and a lower arm, was implemented in the SIMM environment (Delp and Loan,1995). A minimal set of 7 generalized coordinates were chosen to describe the configuration of the shoulder complex (3), elbow (1), and wrist (3). A set of 50 musculotendon units were defined to span each arm (Holzbaur et al.,2005). The 13

Fig. 6. Kinematic parametrization of the upper extremity model. (Left) A set of 7 generalized coordinates were chosen to describe the configuration of the shoulder complex, elbow, and wrist. (Right) A set of 50 musculotendon units were defined to span each arm (Holzbaur et al.,2005).

kinematic parametrization and musculotendon paths are depicted in Fig. 6. Experimental Validation: A set of motion capture experiments were performed to validate the posture-based muscle effort minimization criteria. The subjects performed a set of static tasks designed to isolate upper limb reaching motion. While seated each subject was instructed to pick up a weight and move it to 5 different targets and hold a static configuration at each target for 4 seconds. The posture-based muscle effort criterion (23) was then computed. SIMM was used to generate the maximum muscle induced moments. The results of this analysis showed that the subject’s chosen configuration was typically within several degrees of the predicted configuration associated with minimizing the computed muscle effort (De Sapio et al.,2006). Fig. 7 depicts the results of the muscle effort computations for one of the subject trials with no weight in hand. 3. P ERFORMANCE C HARACTERISTICS IN H UMAN DYNAMIC M OTION This section introduces an extended methodology for identifying physiological characteristics that shape human movement. For this purpose, previously explained human effort minimization strategy (21) was generalized for dynamic skills. As an example illustrating a dynamic skill, we characterized the throwing motion of a football player. In our approach, the performance that suits the footballer can be defined as the ability to achieve maximal ball velocity given the 14

Fig. 7.

Muscle effort variation for one of the subject trials with no weight in hand. Each plot depicts the muscle effort for one

of the five target configurations. The locations of the subject’s chosen configurations are depicted with a ”+”. The full range of motion is depicted by the black silhouettes (±90 from nominal) (De Sapio et al.,2006).

physiological constraints of the system (i.e. limb length, joint range of motion, and muscle strength and contraction velocity). The physiological constraints that affect human motion include the joint constraints (the range of motion at a joint), the segment constraints (the lengths of each segment) and the muscle constraints including physiological cross-section of a muscle, maximum contraction velocity, moment arm and line of action. The football throwing motion was recorded using an 8-camera Vicon motion capture system (OMG plc, Oxford UK) at a capture rate of 120 Hz and the simulation was generated in OpenSim (Delp et al.,2007) and SAI (Khatib et al.,2002) frameworks for the analysis. In order to investigate the muscular effort in dynamic skills in terms of the musculoskeletal parameters, the equation (21) can be written in the form, E = F T Φ(q)F,



Fig. 8. Task-based analysis of muscular effort. The muscle effort variation for selected 5 configurations during a throwing motion represented by ellipsoid expansion model (Khatib and Burdick,1987). The throwing hand trajectory follows the direction of the ellipsoid associated with the minimum instantaneous effort.

where F represents the task requirements and, Φ(q) , J(LT Nc2 L)−1 J T .


Here, the function Φ(q) captures the spacial characterization of the muscular effort measure by connecting the muscle physiology to the resulting task, F , through the Jacobian, J. We studied the dynamic performance characterization and used a graphical representation of the muscular effort function (25) by computing its eigenvalue and eigenvector at a given configuration. The ellipsoids corresponding to the muscular effort were calculated in SAI using the ellipsoid expansion model (Khatib and Burdick,1987). Results of this analysis showed that the direction of the function (25) was minimized in space, which was equivalent to the minimization of the instantaneous effort in the performance of the throwing skill. Fig. 8 shows the ellipsoids corresponding to task-based muscle effort calculations for selected 5 configurations. The current approach involves scaling a musculoskeletal model to match an individuals’ anthropometry and provide subject-specific muscle-tendon and joint parameters, such as; muscle-tendon lengths, moment arms, lines of action, and joint topology. As such, this technique 16

Fig. 9.

(a) The bounds on the joint torques of a multi-degrees of freedom manipulator are mapped to the bounds on the resulting

accelerations to evaluate its dynamic behavior (Khatib and Burdick,1987). (b) The acceleration boundaries of the wrist of a 6 degrees of freedom robotic system: Puma 560. The 3-D parallelepipeds represent the feasible sets of operational space accelerations for different end-effector configurations.

inherently accounts for differences between individuals due to changes in body size. The technique would therefore predict that subjects of different stature would perform the same task (i.e. maximum velocity throwing) with slightly different joint kinematics. Dynamic characterization of human performance needs also to include the analysis of the operational space accelerations. This is motivated by the successful extension of operational space control to analyze the dynamic performance of robotic systems (Khatib and Burdick,1987). In this framework, the idea is to map the analysis of bounds on joint torques to the resulting end-effector accelerations in the workspace of the manipulator (see Fig. 9). Similar model can be applied to characterize and analyze human dynamic skills shaped by the skeletal mechanics as well as the physiological parameters. For this system of n equations and r muscles, Γ is the r × 1 vector of muscle induced joint torques and A is the n × n mass matrix. Using the operational space acceleration/muscle force relationship, x¨ = J(q)A(q)−1 (Γ − g(q) − b(q, q)), ˙


where g(q) is the r × 1 gravity torque vector. The feasible range of accelerations can be determined using (26) given the bounds on the


Fig. 10.

The feasible set of operational space accelerations bounded by the subject’s physiological constraints for selected 5

configurations of a football throwing motion. Accelerations of the throwing hand can be used for the performance characterization of human dynamic skills.

muscle induced torque capacities by, 0 < Γ < LT mmax .


The musculoskeletal model was implemented in SAI control and simulation environment (Khatib et al.,2002) which provided the position Jacobian, J, the muscle Jacobian, L, as well as the muscle induced torques capacities, Γ, and the feasible set of operational accelerations, x¨, for a given configuration, q. The bounds on the feasible set of acceleration were calculated by the convex hull of the affine transformation of a hypercube for r muscles. The hypercube describing the set of allowable muscle induced torques has 2r vertices. Fig.10 illustrates the feasible set of accelerations produced by 12 muscles that contributed most to the resulting acceleration of the hand for selected 5 configurations.


4. D ISCUSSION A robotics-based approach for the synthesis of human motion using task-level control was presented. Concurrent tools in biomechanics and robotics communities enabled our effort to explore natural human motion having benefits in rehabilitation and facilitating development of human-inspired robots. For this purpose, our existing robotic tools were applied to reconstruct and analyze human skills, introducing the approaches of direct marker control, muscle effort criteria and dynamic characterization of human performance. For human motion reconstruction, an extension of operational space control (Khatib,1987) to account for the marker space from human motion capture was presented. The direct marker control algorithm was tested by reconstructing a sequence of motions of a tai-chi master and extracting the joint angles in real-time. This algorithm which currently assumes rigid body dynamics will be extended to account for elastic body links in order to better match subject specific antropometry. For muscular effort minimization, a new posture-based muscle effort criterion was implemented. This criterion is a generalization of the joint decoupled measure used previously (Khatib et al.,2004). The new criterion properly accounts for the cross-joint coupling associated with multi-articular muscle routing kinematics. Through a set of subject trials good correlation between natural reaching postures and those predicted by our posture-based muscle effort criterion were shown. What distinguishes our muscle effort criterion is not its implementation as a posture-based or trajectory-based model, since it is amenable to both, but that it characterizes effort expenditure in terms of musculoskeletal parameters, rather than just skeletal parameters. For the characterization of effort expenditure in terms of both skeletal parameters and the muscle physiology, the muscle effort criterion was implemented to analyze a throwing motion of a football player. The results showed that during the performance of the motion the subject tends to minimize the muscular effort defined by the combination of the force generating kinetics of the muscles as well as the


mechanical advantage, as determined by the muscle routing kinematics and limb mechanics. Additionally, available set of the operational space accelerations of the throwing hand were used to support the characterization of the same dynamic skill. One might expect localized muscle fatigue or differences in muscle strength to alter movement patterns when performing the same task. Muscle fatigue, atrophy, or strength can be simulated within our muscular effort criteria by altering the muscle parameters within the model (muscle capacity, c). In the case of fatigue, for example, the force producing capacity of a muscle group can be altered, and the generalized approach would predict a slightly different movement trajectory to compensate for this reduced capacity. As such, the methods presented are generalized and not limited solely to optimal movements. Accurate modeling and detailed understanding of human motion will have a significant impact on a host of domains: from the rehabilitation of patients with physical impairments to the training of athletes or the design of machines for physical therapy and sport. In the case of rehabilitation, a patient would benefit from knowing what movement pattern might influence loads on a specific joint or tissue. For example, a patient who has undergone arthroscopic knee meniscectomy is at high risk of developing knee joint osteoarthritis, particularly if they walk with large knee adduction (varus) moments. In this scenario, the patient would benefit from knowing what movement pattern could be used to reduce loading on the medial compartment of the knee during walking, thus alleviating the stresses on the articular surface of the knee and reducing the risk of developing osteoarthritis. An additional term describing the loads at the knee could easily be added to the current optimization criteria and the generalized robotics technique could be used to predict a novel gait pattern for the patient that minimizes energy expenditure during walking as well as reducing the loads on the knee. The patient would then be taught this new gait pattern using visual or haptic feedback. This scenario is currently being investigated by the authors. In spite of the great complexity of natural human motion, the robotic-based analysis of human performance provides substantial benefits to researchers focused on restoring or improving


human movement. Human motor performance depends on skilled motor coordination as well as physical strength. Optimal movements such as those exhibited by highly skilled practitioners in sports and the martial arts provide inspiration for developers of humanoid robots. This dual dependency motivates our work on the analysis and synthesis of human motion. 5. ACKNOWLEDGEMENT The financial support of the Simbios National Center for Biomedical Computing Grant (, NIH GM072970) and KAUST (King Abdullah University of Science and Technology) are gratefully acknowledged. R EFERENCES [Choi and Ko,1999] Choi, K., and Ko, H.. (1999). On-line Motion Retargetting. Seventh Pacific Conference on Computer Graphics and Applications, pp. 32. [De Sapio et al.,2005] De Sapio, V., Warren, J., Khatib, O., and Delp, S. (2005). Simulating the Task-level Control of Human Motion: A Methodology and Framework for Implementation. The Visual Computer, vol. 21, no. 5, pp. 289-302. [De Sapio et al.,2005] De Sapio, V., Holzbaur, K. R., and Khatib, O. (2005). The Control of Kinematically Constrained Shoulder Complexes: Physiological and Humanoid Examples. Proceedings of the 2005 IEEE International Conference on Robotics and Automation, vol. 1–10, pp. 2952–2959, Barcelona. [De Sapio et al.,2006] De Sapio, V., Warren, J., Khatib, O. (2006). Predicting Reaching Postures Using a Kinematically Constrained Shoulder Model. In J. Lenarcic and B. Roth, eds. Advances in Robot Kinematics, pp. 209-218, Springer. [Delp et al.,1990] Delp, S.L., Loan, J.P., Hoy, M.G., Zajac, F.E., Topp E.L., Rosen, J.M. (1990) An interactive graphics-based model of the lower extremity to study orthopaedic surgical procedures. IEEE Transactions on Biomedical Engineering, vol. 37, pp. 757-767. [Delp and Loan,1995] Delp, S., and Loan, P. (1995) A software system to develop and analyse models of musculoskeletal structures. Computers in Biology and Medicine, vol. 25, pp. 21–34. 21

[Delp et al.,2007] Delp, S.L., Anderson, F.C., Arnold, A.S., Loan, P., Habib, A., John, C.T., Guendelman, E., Thelan, D.G. (2007). OpenSim: Open-source software to create and analyze dynamic simulations of movement. IEEE Transactions on Biomedical Engineering , vol. 55, pp. 1940-1950. [Demircan et al.,2008] Demircan, E., Sentis, L., De Sapio, V., and Khatib, O. (2008). Human Motion Reconstruction by Direct Control of Marker Trajectories. In J. Lenarcic and P. Wenger, eds. Advances in Robot Kinematis, pp. 263-272, Springer Netherlands. [Dempster,1955] Dempster, W. T. (1955). Space requirements of the seated operator. WADC Technical Report, pp. 55-159. [Grochow et al.,2004] Grochow, K., Martin, S. L., Hertzmann, A., Popovic, Z. (2004). Style-based inverse kinematics. ACM Transactions on Graphics (TOG), Proceedings of the 2004 SIGGRAPH Conference, vol. 23, no. 3, pp. 522–531. [Hermens and Gielen,2004] Hermens F., Gielen S. (2004), Posture-based or trajectory-based movement planning: a comparison of direct and indirect pointing movements, Experimental Brain Research, vol. 159, no. 3, pp. 340–348. [Holzbaur et al.,2005] Holzbaur, K.R., Murray, W.M., and Delp, S.L. (2005). A Model of the Upper Extremity for Simulating Musculoskeletal Surgery and Analyzing Neuromuscular Control. Annals of Biomedical Engineering, vol. 33, no. 6, pp. 829-840. [Kang et al.,2005] Kang, T., He, J., Helms Tillery, S.I. (2005) Determining natural arm configuration along a reaching trajectory. Experimental Brain Research, vol. 167, no. 3, pp. 352–361. [Khatib and Le Maitre,1978] Khatib, O., Le Maitre, J.F. (1978) Dynamic Control of Manipulators Operating in a Complex Environment, Proceedings RoManSy’78, 3rd CISM-IFToMM Symp. Theory and Practice of Robots and Manipulators, Udine, Italy, pp. 267-282. [Khatib,1986] Khatib, O. (1986) Real-Time Obstacle Avoidance for Manipulators and Mobile Robots. International Journal of Robotic Research, vol. 5, no. 1, pp. 90–98.


[Khatib,1987] Khatib, O. (1987). A unified approach for motion and force control of robot manipulators: The operational space formulation. International Journal of Robotics and Automation, vol. 3(1), pp. 43-53. [Khatib and Burdick,1987] Khatib, O. and Burdick, J. (1987). Optimization of Dynamics in Manipulator Design: The Operational Space Formulation. International Journal of Robotics and Automation, vol. 2(2), pp. 90-98. [Khatib,1995] Khatib, O. (1995). Inertial properties in robotic manipulation: An object level framework. International Journal of Robotics Research, vol. 14, no. 1, pp. 19–36. [Khatib et al.,1999] Khatib, O., Yokoi, K., Brock, O., Chang, K.-S., and Casal, A. (1999). Robots in human environments: Basic autonomous capabilities. Int. J. Robotics Research, vol. 18(7), pp. 175–183. [Khatib et al.,2002] Khatib, O., Brock, O., Chang, K., Conti, F., Ruspini, D., Sentis, L. (2002). Robotics and Interactive Simulation. Communications of the ACM, vol. 45, no. 3, pp. 46–51. [Khatib et al.,2004] Khatib, O., Sentis, L., Park, J., Warren, J. (2004). Whole-Body Dynamic Behavior and Control of Human-Like Robots. International Journal of Humanoid Robotics, vol. 1, no. 1, pp. 29–43. [Khatib et al.,2004] Khatib, O., Warren, J., De Sapio, V., and Sentis, L. (2004). Human-like motion from physiologically-based potential energies. In Lenarˇciˇc J, Galletti C, eds. On advances in robot kinematics, pp. 149–163, Kluwer. [Khatib et al.,2008] Khatib, O., Sentis, L., and Park, J. (2008). A Unified Framework for Whole-Body Humanoid Robot Control With Multiple Constraints and Contacts. European Robotics Symposium, Springer Tracts in Advanced Robotics - STAR. [Lacquaniti and Soechting,1982] Lacquaniti, F., Soechting, J.F. (1982). Coordination of Arm and Wrist Motion During a Reaching Task. Journal of Neuroscience, vol. 2, no. 4, pp. 399-408. [Lee and Shin,1999] Lee J., Shin S.Y. (1999) A hierarchical approach to interactive motion editing for human-like figures. Proceedings of the 26th Annual Conference on Computer


Graphics and Interactive Techniques, pp. 39–48. [Nakamura et al.,2003] Nakamura Y., Yamane K., Suzuki I., and Fujita Y. (2003). Dynamic Computation of Musculo-Skeletal Human Model Based on Efficient Algorithm for Closed Kinematic Chains, Proceedings of the 2nd International Symposium on Adaptive Motion of Animals and Machines, Kyoto, Japan, SaP-I-2. [Savenko and Clapworthy,2002] Savenko, A., Clapworthy, G. (2002). Using Motion Analysis Techniques for Motion Retargetting. Sixth International Conference on Information Visualization, iv, pp. 110. [Sentis and Khatib,2005] Sentis, L. and Khatib, O. (2005). Synthesis of Whole-Body Behaviors Through Hierarchical Control of Behavioral Primitives. International Journal of Humanoid Robotics, vol. 2(4), pp. 505–518. [Soechting et al.,1995] Soechting, J.F., Buneo, C.A., Herrmann, U., Flanders, M. (1995). Moving Effortlessly in Three Dimensions: Does Donders Law Apply to Arm Movement?, Journal of Neuroscience, vol. 15, no. 9, pp. 6271–6280. [Thelen et al.,2003] Thelen, D.G., Anderson, F.C., and Delp, S.L. (2003). Generating dynamic simulations of movement using computed muscle control. Journal of Biomechanics, vol. 36, pp. 321-328. [Uno et al.,1989] Uno, Y., Kawato, M., Suzuki, R. (1989). Formation and control of optimal trajectory in human multijoint arm movement. Biological Cybernetics, vol. 61, pp. 89–101. [Vetter et al.,2002] Vetter P., Flash T., Wolpert D.M. (2002). Planning Movements in a Simple Redundant Task, Current Biology, vol. 12, no. 6, pp. 488–491. [Zajac,1989] Zajac, F.E. (1989). Muscle and tendon: properties, models, scaling, and application to biomechanics and motor control. Critical Reviews in Biomedical Engineering, vol. 17, no. 4, pp. 359–411. [Zajac,1993] Zajac, F.E. (1993). Muscle coordination of movement: a perspective. Journal of Biomechanics, vol. 26, pp. 109–124.



Robotics-based Synthesis of Human Motion - CiteSeerX

Robotics-based Synthesis of Human Motion Khatib, O.a , Demircan, E.a , De Sapio, V.a,b , Sentis, L.a , Besier, T.c , Delp, S.d a Artificial Intellige...

2MB Sizes 1 Downloads 20 Views

Recommend Documents

Human Hand Motion Analysis and Synthesis of Optimal Power Grasps
Jump to Bio-inspired Grasping Algorithm for a Robotic Hand - 3 and obtained from the analysis of the human hand behaviou

A Research Synthesis - CiteSeerX
William R. Penuel. SRI International. Abstract ... Education found just 19 studies that had analyzed outcomes (Penuel et

Synthesis, processing and material properties of - CiteSeerX
by the low stability towards atmospheric oxidation and the lack of processability often characterizing these materials.

Design and Optimization of Human Motion and Motion Aid Increasing
Amenity space design for comfortable life. Dynamics for Human. Robot for Human. Environment for Human. 4. Collaboration.

Human Relations - CiteSeerX
We propose to extend the Exit, Voice, Loyalty, and Neglect (EVLN) model of employees' responses to adverse organizationa

Generating Natural Motion in an Android by Mapping Human Motion
keywords: Learning control systems, motion analysis, humanlike motion, human-robot imitation, android science, appearanc

Imitating Human Dance Motion through Motion Structure Analysis
Abstract. This paper presents the method for importing human dance motion into humanoid robots through visual obser- vat

Design, Synthesis, Antinociceptive and Anti-Inflammatory - CiteSeerX
Nov 28, 2012 - Luciano Morais Lião 6, Lídia Moreira Lima 1,2 and Eliezer J. Barreiro 1,2,*. 1 LASSBio® ... 2 Pós-graduaç

Generating Natural Motion in an Android by Mapping Human Motion
which joint angles would make the robot's posture appear similar to the human performer; and (2) the motion that we perc

synthesis of design, testing and analysis research on steel - CiteSeerX
Oct 1, 2005 - Synthesis of Research on Steel Column Base Plate Connections .... Investigation of Column Base Design Para