STABILITY ANALYSIS AND CONTROL DESIGN

Transcription

1 STABILITY ANALYSIS AND CONTROL DESIGN WITH IMPACTS AND FRICTION Michael Posa Massachusetts Institute of Technology Collaboration with Mark Tobenkin and Russ Tedrake ICRA Workshop on Robust Optimization-Based Control and Planning for Legged Robots May 16, 2016

2 CONTACT AND CONTROL Complex contact scenarios make typical control policies very brittle Require fast and accurate sensing of contact state Struggle when unexpected contact occurs Carefully tuned, ad-hoc approaches near contact events

3 CONTACT AND CONTROL Complex contact scenarios make typical control policies very brittle Require fast and accurate sensing of contact state Struggle when unexpected contact occurs Carefully tuned, ad-hoc approaches near contact events Algorithms for finding and verifying controllers robust to contact dynamics?

4 WHAT MAKES CONTACT HARD?

5 WHAT MAKES CONTACT HARD? It's where our intuitive designs struggle

6 Good techniques for ODEs: WHAT MAKES CONTACT HARD? It's where our intuitive designs struggle

7 WHAT MAKES CONTACT HARD? Good techniques for ODEs: Discontinuities, but... Continuity of solutions w.r.t initial conditions It's where our intuitive designs struggle

8 WHAT MAKES CONTACT HARD? Good techniques for ODEs: Discontinuities, but... Continuity of solutions w.r.t initial conditions Uniqueness It's where our intuitive designs struggle

9 WHAT MAKES CONTACT HARD? Good techniques for ODEs: Discontinuities, but... Continuity of solutions w.r.t initial conditions Uniqueness Zeno Phenomena It's where our intuitive designs struggle

10 MODELS: DYNAMICS, IMPACTS, AND FRICTION Joint coordinates and velocities Rigid-body dynamics: Contact gap function and Jacobian Impulsive, inelastic impacts "Incremental" frictional impacts [Routh 1891] Coulomb friction

11 VERIFICATION AND ANALYSIS Lyapunov functions powerful for smooth systems Tools from convex optimization: sums-of-squares (SOS) [Parrilo 2000, Lasserre 2001] Hybrid versions analyze every contact mode and transition Exponential in number of contacts [Papachristodoulou 2009] Instead, use Measure Differential Inclusion (MDI) framework [Moreau 1988, Stewart 2000, and others] Exploit complementarity struction Pose tractable SOS programs

12 LYAPUNOV FUNCTIONS Capture global or regional stability properties of nonlinear systems Example: reverse-time Van der Pol oscillator [Tan and Packard]

13 SUMS-OF-SQUARES AND LYAPUNOV FUNCTIONS Fundamentally questions of non-negativity If a polynomial, is NP-hard Replace with a sufficient condition Finding is a convex constraint in a Semidefinite Program (SDP) * [Parillo 2000, Lasserre 2001] For some basis (e.g. monomials up to degree ) * Generalization of linear programming

14 S-PROCEDURE Regional analysis for semialgebraic set For example, Sufficient condition: -sublevel set of in blue

15 5 state model 2 contact points Exhibits Zeno phenomena PASSIVE RIMLESS WHEEL Goal: verify stable region [Posa, Tobenkin, and Tedrake (TAC 2016)], [Posa, Tobenkin, and Tedrake (HSCC 2013)]

16 LYAPUNOV FUNCTIONS FOR MDIS Measure differential inclusion modeling framework, instead of ODE Encompasses discontinuities and non-uniqueness continuous and of locally bounded variation Dynamics from set-valued functions Simulation requires selection from set For verification, take permissive view Prove for all possible scenarios

17 HOW TO EFFICIENTLY VERIFY? admissable

18 HOW TO EFFICIENTLY VERIFY? admissable

19 HOW TO EFFICIENTLY VERIFY? admissable Observation: sufficient to check

20 LEVERAGING STRUCTURE What are admissable?

21 LEVERAGING STRUCTURE What are admissable? Leverage complementarity formulation Non-penetration: Friction cone: Contact mode logic: Semialgebraic in

22 LEVERAGING STRUCTURE What are admissable?

23 TRACTABLE OPTIMIZATION PROGRAM Theorem [TAC 2016]: For contacts, can eliminate normal force variables and verify contacts independently Proof exploits continuity and homogeneity in Reduced conditions:

24 TRACTABLE OPTIMIZATION PROGRAM For -dimensional state, contacts, and polynomial degree, Hybrid * MDI multiple Lyapunov functions single Lyapunov function SOS constraints variables * Exact count depends on formulation of sliding and sticking contact

25 CONTROL DESIGN AND VERIFICATION 7 state model 2 contact points 1 input Control design bilinear in and Solve via alternating sequence of SOS programs θ Φ What is possible with a continuous feedback policy? Goal: maximize stable region z x

26 CONTROL DESIGN AND VERIFICATION Sampled 2D slice of state space and simulate Verified region in blue Stable samples as red dots

27 UNSAFE REGION AVOIDANCE

28 DISCUSSION A step toward control and verification in complex contact scenarios Restricted to fairly simple models Beyond capability of hybrid formulation SOS/SDP relatively immature Introduce model/terrain uncertainty at computational cost Preliminary work in exploiting simple models for push recovery Relevant publications Posa, Tobenkin, and Tedrake. Lyapunov analysis of rigid body systems with impacts and friction via sums-of-squares. HSCC, Best Paper Award. Posa, Tobenkin, and Tedrake. Stability analysis and control of rigid-body systems with impacts and friction. TAC, 2016.