CoM state-space cascading manifolds for planning dynamic walking in very rough terrain

Luis Sentis* and Benito Fernandez** lsentis@austin.utexas.edu, benito@mail.utexas.edu The University of Texas in Austin; 1 University Station C2200, Austin, TX 78712, USA

1 Introduction Formulating rough terrain locomotion implies dealing with nonlinear/nonintegrable models of Center of Mass (CoM) behavior with respect to the contact state. Our hypothesis is that center of mass phase curves can be created using perturbation theory and then used to nd contact transitions to produce dynamic walking in very rough terrains. This extended abstract aims at validating this hypothesis. In dynamic walking we can classify techniques in various categories: (1) trajectory-based techniques, (2) limit cyclebased techniques, (3) prediction of contact, and (4) hybrids of the previous three. Trajectory-based techniques: These are techniques that track a time-based joint or task space trajectory according to some locomotion model such as the Zero Moment Point (ZMP). [6] and [8] represent some of the latest developments in this area. Prediction of contact placement: These are techniques that use dynamics to estimate suitable contact transitions. In [14], simple dynamic models are used to predict the placement of next contacts to achieve desire gait patterns.Some of the most recent work can be found in in [2, 4, 13]. Limit cycle based techniques: McGeer [11] pioneered the eld of passive dynamic walking. In [5] the authors study orbital stability, and the effect of feedback control to achieve asymptotic stability. Optimization of open-loop stability is investigated in [12]. In [15], the authors analyze the energetic cost of bipedal walking and running. In [19], the authors developed a dynamic walker using articial muscles. In [20], a methodology for the analysis of state-space behavior and feedback control are presented. Step recovery in response to perturbations is studied in [18]. In [16], the interplay between robustness against perturbations and leg compliance is investigated. Hybrid methods: In [21], the stability of passive walkers is studied. Stochastic models of stability and its application for walking on moderately rough unmodeled terrain are studied in [3]. The design of non-periodic locomotion for uneven terrain is investigated in [10]. In [17], the authors explore the design of pasitivity-based controllers to achieve walking on different ground slopes. Optimization-based techniques for locomotion in rough terrains are presented in [22]. In [7], the authors exploit optimization as a means to plan locomotion. Figure 1: Data extraction from human walk: A human subject walks over a rough terrain. Marker tracking is implemented and used to extract approximate CoM paths as well as Sagittal and vertical CoM trajectories and velocities. 2 Summary of our approach We tackle here rough terrain locomotion by exploring CoM state-space manifolds and transitional contact states. The approach is applicable to all terrains with special emphasis to very rough environments. Our approach, can be explained algorithmically in terms of various phases, namely (1) geometric planning, (2) perturbation-based CoM phase generation, and (3) step solver based on adjacent phase curves The geometric planning phase consists on applying kinematic planning techniques to obtain initial guesses of feet contact locations and CoM geometric path. Perturbation-based CoM phase generation is our rst contribution and consists on: (1) formulating CoM accelerations based on the contact state, (2) include the dependency between Sagittal and vertical accelerations due to the desired CoM geometric path, and (3) use perturbation theory to obtain phase curves of the CoM in the neighborhood of the step contact. The step solver, is our second contribution and consists on tting polynomials to phase curves of adjacent steps and nding the roots of the differential polynomial between inmediate neighbors. 3 Perturbation-based CoM phase generation We present here a solution for robot locomotion in the Sagittal/vertical plane. Using a human-size robot model, we consider a variable stepped terrain with height variations between 40 [cm] and width variations between 30-40 [cm]. The goal of the planner is to maneuver the robot through the total length of the terrain. The speed specications are given to cruise the terrain at an average velocity of 0.6 [m/s], al-

though this choice could be arbitrary. We also assume that the robot starts and nishes with zero velocities and it increases velocity through the steps according to a trapezoidal prole. Velocity specications are given only at each new step, corresponding to the moment when the center of mass Sagittal position crosses the corresponding supporting foot, namely the apex of the step. As such, they are equivalent to boundary conditions to solve the state space behavior. Lets consider steps to be spanned from apex to apex. Also for simplicity, we consider only single-support phases, with instantaneous transition between feet. We also assume that the contact locations and the geometric path of the center of mass are given by a kinematic planner, and we assume a point mass model of the robot, with all the weight located at its waist. For every contact state, we formulate dynamic equilibrium of moments, i.e. contact reactions vs. inertial and gravitational moments. This relationship yields a well-known solution that relates Sagittal accelerations with respect to center of mass Sagittal and vertical distances to contact locations, and with the latter multiplied by vertical accelerations plus gravity. CoM moments do not appear due to point mass conditions. Most researchers simplify the above equation by assuming xed vertical CoM and feet conditions. However, to walk in very rough uneven terrains this assumption is no longer valid. Instead, we assume unconstrained vertical CoM and feet variations, assuming they will be kinematically feasible. We seek to nd a state-space manifold of CoM behavior due to the varying contact conditions and desired CoM kinematic path. We refer to perturbation theory [1, 9] to address the difculty of solving nonintegrable equations. In particular, perturbation theory, has been widely used to solve the trajectory of celestial bodies and pendulums. Perturbation theory, is a set of methods that enable to approximate solutions from problems that dont have exact solutions, by looking into the solution of an exact related problem. In our case, we have the exact solution of accelerations given positions and pressure points and we seek to approximate the solution of the position versus the velocity, i.e. the state-space trajectory. Using perturbation theory, we obtain the incremental relationship between CoM positions and velocities for each contact state, thus yielding state-space CoM specications. Because we operate in state-space we remove time as a variable. The CoM manifolds, by construction, describe the CoM behavior before and after each apex. If we combine neighboring manifolds, the contact transition state can be determined nding the interseccion of the curves. This approach is the key contribution of this work (see Figure 2). Using the prescribed CoM kinematic path, it is now possible to extract the corresponding CoM manifold in the vertical direction. Moreover, given the contact transition states it is also possible to derive feet state-space curves. This information in turn, can be utilized to create joint velocity or torque feedback controllers to make the CoM manifold an attractor.

Figure 2: Concatenation of steps: The top graph shows the kinematic trajectory of the human CoM (see Figure 1 for the extraction of motion capture data) versus a piecewise linear approximation that we use to generate the automatic walking simulation. The red dots correspond to the position of the foot contacts. The bottom gure shows Matlab plots of Sagittal phase curves for the human and the automatic simulation. The red circles correspond to apexes of the steps. The green squares correspond to contact transitions of the automatic walk. The blue squares correspond to contact transitions of the human walk. Notice, that during the climbing of the rst step of the stairs results in a smooth CoM pattern for the human walk. This is due to the smoothening effect of dual contact during the stance phase. This is not the case during the automatic walk because we have neglected the stance phase and therefore the transitions between contacts are instantaneous. Besides this difference, the rest of the walk correlates well. 4 Open questions For this seminal work, we would like to address two of these issues: robustness and the role of internal forces during multicontact phases. Robustness issues are critical to the success of implementing dynamic locomotion. We can study the stability robustness of a specic manifold to parameter uncertainty or to external disturbances. That is, how much parametric uncertainty is allowed before the CoM manifold is no longer attractive. The least-known parameters are friction forces coefcients in the joints and with the ground. By looking at the parameter region around the estimated values, this technique can be used to determine if and where the algorithm would fail to obtain a suitable transitional state. We could also address the effect of joint compliance in the stability robustness of the system. Furthermore, the proposed approach can be used to design the stiffness requirements that guarantee disturbance rejection (maintaining stability robustness) while traversing rough terrains. Locomotion under multi-contact states provides advantages in the decision-making scenario on how and where to determine the transition state. Rather than a single phase manifold, multi-contact conditions lead to manifolds where many solutions are plausible. As a result, the range of motions is increased and optimization algorithms can be carried out to maximize robustnes, minimize energy consumption, etc. We plan to discuss our take on this problem and propose potential extensions of the algorithm to multi-contact scenarios.

References [1] C. Bender and S. Orszag. Advanced Mathematics for Scientists and Engineers; Asympotic Methods and Perturbation Theory. McGraw Hill, 1991. [2] T. Bretl and S. Lall. Testing static equilibrium for legged robots. IEEE Transactions on Robotics, 24(4):794807, August 2008. [3] K. Byl and R. Tedrake. Metastable walking machines. The International Journal of Robotics Research, 28(8):10401064, 2009. [4] C. Collette, A. Micaelli, C. Andriot, and P. Lemerle. Robust balance optimization control of humanoid robots with multiple non coplanar grasps and frictional contacts. In Proceedings of the IEEE International Conference on Robotics and Automation, Pasadena, USA, May 2008. [5] A. Goswami, B. Espiau, and A. Kermane. Limit cycles and their stability in a passive bipedal gait. In Proceedings of the IEEE International Conference on Robotics and Automation, pages 246251, April 1996. [6] K. Harada, S. Kajita, K. Kaneko, and H. Hirukawa. Zmp analysis for arm/leg coordination. In Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 7581, Las Vegas, USA, October 2003. [7] K. Hauser, T. Bretl, K. Harada, and J. Latombe. Using motion primitives in probabilistic sample-based planning for humanoid robots. In Workshop on Algorithmic Foundations of Robotic (WAFR), New York, USA, July 2006. [8] S. Kajita, M. Morisawa, K. Harada, K. Kaneko, F. Kanehiro, K. Fujiwara, and H. Hirukawa. Biped walking pattern generator allowing auxiliary zmp control. In Intelligent Robots and Systems, 2006 IEEE/RSJ International Conference on, pages 29932999, october 2006. [9] H. Khalil. Nonlinear Systems. Prentice-Hall, Inc., 1996.

[10] I. R. Manchester, U. Mettin, F. Iida, and R. Tedrake. Stable dynamic walking over uneven terrain. The International Journal of Robotics Research, 30(3):265279, 2011. [11] T. McGeer. Passive dynamic walking. The International Journal of Robotics Research, 9(2):6268, 1990. [12] K. Mombaur, H. Bock, J. Schloder, and R. Longman. Self-stabilizing somersaults. Robotics, IEEE Transactions on, 21(6):1148 1157, december 2005. [13] J. Pratt and R. Tedrake. Velocity-based stability margins for fast bipedal walking. In M. Diehl and K. Mombaur, editors, Fast Motions in Biomechanics and Robotics, volume 340, pages 299324. 2006. [14] M. Raibert. Legged Robots that Balance. MIT Press, Cambridge, Ma., 1986.

[15] A. Ruina, J. Bertram, and M. Srinivasan. A collisional model of the energetic cost of support work qualitatively explains leg sequencing in walking and galloping, pseudo-elastic leg behavior in running and the walk-to-run transition. Journal of Theoretical Biology, (237):170192, 2005. [16] J. Rummel, Y. Blum, and A. Seyfarth. Robust and efcient walking with spring-like legs. Bioinspiration & Biomimetics, 5(4):046004, 2010. [17] M. Spong, J. Holm, and D. Lee. Passivity-based control of bipedal locomotion. Robotics Automation Magazine, IEEE, 14(2):30 40, june 2007. [18] B. Stephens and C. Atkeson. Modeling and control of periodic humanoid balance using the linear biped model. In Humanoid Robots, 2009. Humanoids 2009. 9th IEEE-RAS International Conference on, pages 379 384, december 2009. [19] T. Takuma and K. Hosoda. Controlling the walking period of a pneumatic muscle walker. The International Journal of Robotics Research, 25(9):861866, 2006. [20] E. Westervelt, J. Grizzle, C. Chevallereau, J. Choi, and B. Morris. Feebdack control of dynamic bipedal robt locomotion. CRC Oress, 2007. [21] M. Wisse, A. Schwab, R. van der Linde, and F. van der Helm. How to keep from falling forward: elementary swing leg action for passive dynamic walkers. Robotics, IEEE Transactions on, 21(3):393 401, june 2005. [22] M. Zucker, J. Bagnell, C. Atkeson, and J. Kuffner. An optimization approach to rough terrain locomotion. In Robotics and Automation (ICRA), 2010 IEEE International Conference on, pages 3589 3595, may 2010.