Discrete Variational Mechanics Benjamin Stephens J.E. Marsden and M. West, “Discrete mechanics and variational integrators,” Acta Numerica, No.
Download ReportTranscript Discrete Variational Mechanics Benjamin Stephens J.E. Marsden and M. West, “Discrete mechanics and variational integrators,” Acta Numerica, No.
Discrete Variational Mechanics Benjamin Stephens J.E. Marsden and M. West, “Discrete mechanics and variational integrators,” Acta Numerica, No. 10, pp. 357-514, 2001 M. West “Variational Integrators,” PhD Thesis, Caltech, 2004 1 About My Research • Humanoid balance using simple models • Compliant floating body force control • Dynamic push recovery planning by trajectory optimization C C(t ) Fˆ PL PL (t ) PR http://www.cs.cmu.edu/~bstephe1 2 http://www.cs.cmu.edu/~bstephe1 3 But this talk is not about that… The Principle of Least Action The spectacle of the universe seems all the more grand and beautiful and worthy of its Author, when one considers that it is all derived from a small number of laws laid down most wisely. -Maupertuis, 1746 5 The Main Idea • Equations of motion are derived from a variational principle • Traditional integrators discretize the equations of motion • Variational integrators discretize the variational principle 6 Motivation • Physically meaningful dynamics simulation Stein A., Desbrun M. “Discrete geometric mechanics for variational time integrators” in 7 Discrete Differential Geometry. ACM SIGGRAPH Course Notes, 2006 Goals for the Talk • Fundamentals (and a little History) • Simple Examples/Comparisons • Related Work and Applications • Discussion 8 The Continuous Lagrangian • Q – configuration space • TQ – tangent (velocity) space • L:TQ→R L(q, q ) T (q, q ) U (q) Lagrangian Kinetic Energy Potential Energy 9 Variation of the Lagrangian • Principle of Least Action = the function, q*(t), minimizes the integral of the Lagrangian T T 0 0 L(q * (t ), q * (t ))dt L(q * (t ) q(t ), q * (t ) q (t ))dt “Calculus of Variations” ~ Lagrange, 1760 T Variation of trajectory with endpoints fixed L( q (t ), q (t )) dt 0 0 “Hamilton’s Principle” ~1835 10 Continuous Lagrangian L d L 0 q dt q “Euler-Lagrange Equations” 11 Continuous Mechanics L(q(t ), q (t )) T (q, q ) U (q) L d L d (T U ) q dt q q dt q T U d T q q dt q T U 2T 2T q q q q qq qq M (q)q C (q, q ) G(q) 0 12 The Discrete Lagrangian T h • L:QxQ→R Lq(t ), q(t )dt L q , q d k k 1 , h T qk 1 qk Ld qk , qk 1 , h L qk , h h L qk h qk 1 13 Variation of Discrete Lagrangian D2 Ld qk 1, qk , t D1Ld qk , qk 1 , t 0 “Discrete Euler-Lagrange Equations” 14 Variational Integrator • Solve for qk 1: D2 Ld qk 1, qk , h D1Ld qk , qk 1, h 0 qk qk qk 1 h L qk 1 , h qk qk 1 qk L qk , h qk qk 1 L qk qk 1 L qk 1 qk L q , h q , k 1 k 1 qk , q h q h h q h 0 L qk 1 qk h qk , h q 0 15 Solution: Nonlinear Root Finder f (qk 1 ) D2 Ld qk 1, qk , h D1Ld qk , qk 1 , h 0 i 1 k 1 q q i k 1 i k 1 i k 1 f (q ) Df (q ) 16 Simple Example: Spring-Mass 1 2 1 2 Lq, q mx kx 2 2 • Continuous Lagrangian: • Euler-Lagrange Equations: L d L kx mx 0 q dt q • Simple Integration Scheme: 1 2 k xk 1 xk hx k h xk 2 m k xk 1 x k h xk m 17 Simple Example: Spring-Mass • Discrete Lagrangian: 1 xk 1 xk 1 2 Ld xk , xk 1 , h m kxk 2 h 2 2 • Discrete Euler-Lagrange Equations: D2 Ld xk 1, xk , h D1Ld xk , xk 1 , h 0 m 2 xk 1 2 xk xk 1 kx k 0 h • Integration: 2 h k xk xk 1 xk 1 2 m 18 Comparison: 3 Types of Integrators • Euler – easiest, least accurate • Runge-Kutta – more complicated, more accurate • Variational – EASY & ACCURATE! 19 Euler h=0.001 Runge-Kutta (ode45) Variational h=0.001 0.04 0.02 velocity 0 -0.02 -0.04 -0.06 -0.08 0.97 0.98 0.99 1 1.01 1.02 position 1.03 1.04 1.05 1.06 20 0.51 Euler h=0.001 Runge-Kutta (ode45) Variational h=0.001 0.508 Energy 0.506 0.504 0.502 0.5 0.498 Notice: 0 10 20 30 40 50 time (s) 60 70 80 90 100 •Energy does not dissipate over time •Energy error is bounded 21 Variational Integrators are “Symplectic” • Simple explanation: area of the cat head remains constant over time Stein A., Desbrun M. “Discrete geometric mechanics for variational time integrators” in 22 Discrete Differential Geometry. ACM SIGGRAPH Course Notes, 2006 Forcing Functions • Discretization of Lagrange–d’Alembert principle 23 Constraints D2 Ld qk 1 , qk , h D1Ld qk , qk 1 , h g (qk )T k f ( zk 1 ) 0 g (qk 1 ) z k 1 z i 1 k 1 z qk 1 k i k 1 i k 1 i k 1 f (z ) Df ( z ) 24 Example: Constrained Double Pendulum w/ Damping 2 x g ( q) 0 y x y q 1 2 1 0 0 F (q) K1 K 2 ( x, y ) 25 Example: Constrained Double Pendulum w/ Damping • Constraints strictly enforced, h=0.1 No stabilization heuristics required! 26 Complex Examples From Literature E. Johnson, T. Murphey, “Scalable Variational Integrators for Constrained Mechanical Systems in Generalized Coordinates,” IEEE Transactions on Robotics, 2009 a.k.a “Beware of ODE” 27 Complex Examples From Literature Variational Integrator ODE 28 Complex Examples From Literature 29 log Complex Examples From Literature Timestep was decreased until error was below threshold, leading to longer runtimes. 30 Applications • Marionette Robots E. Johnson and T. Murphey, “Discrete and Continuous Mechanics for Tree Represenatations of Mechanical Systems,” ICRA 2008 31 Applications • Hand modeling E. Johnson, K. Morris and T. Murphey, “A Variational Approach to Stand-Based Modeling 32 of the Human Hand,” Algorithmic Foundations of Robotics VII, 2009 Applications • Non-smooth dynamics Fetecau, R. C. and Marsden, J. E. and Ortiz, M. and West, M. (2003) Nonsmooth Lagrangian mechanics and variational collision integrators. SIAM Journal on Applied Dynamical Systems 33 Applications • Structural Mechanics Kedar G. Kale and Adrian J. Lew, “Parallel asynchronous variational integrators,” International Journal for Numerical Methods in Engineering, 2007 34 Applications • Trajectory optimization O. Junge, J.E. Marsden, S. Ober-Blöbaum, “Discrete Mechanics and Optimal Control”, in Proccedings of the 16th IFAC World Congress, 2005 35 Summary • Discretization of the variational principle results in symplectic discrete equations of motion • Variational integrators perform better than almost all other integrators. • This work is being applied to the analysis of robotic systems 36 Discussion • What else can this idea be applied to? – Optimal Control is also derived from a variational principle (“Pontryagin’s Minimum Principle”). • This idea should be taught in calculus and/or dynamics courses. • We don’t need accurate simulation because real systems never agree. 37