6/26/2015 Pop Quiz for Tony Bipedal Locomotion on Small Feet Jessy Grizzle Elmer G. Gilbert Distinguished University Professor Levin Professor of Engineering ECE and ME Departments Can you give the first name of either D’Alembert or Lagrange? Jean le Rond d'Alembert and Joseph Louis Lagrange What is the air-speed velocity of an unladen swallow? What do you mean? An African or European swallow? Whose birthday am I unable to celebrate today? My wife’s birthday! Major Themes Bipedal Locomotion on Small Feet Jessy Grizzle Elmer G. Gilbert Distinguished University Professor Levin Professor of Engineering ECE and ME Departments From Simple More Complex Walking may be easy for people, still challenging for robots! Control of balance on small feet enables agility and robustness on normal feet. The unreasonable effectiveness of geometry + nonlinear control + parameter optimization. From Not So Simple Very Complex Bipedal robots @ Aaron Ames 1 6/26/2015 MABEL: Planar & Underactuated Jonathan Hurst Does the bar hold up the robot? Control Algorithms Large Springs K. Sreenath H. W. Park 1m 65 Kg Bipedal robots outdoors J. Hurst: 3 ATRIAS robots built & delivered 11 Ultimate Goal: MARLO Lateral leg motion MARLO Carnegie Mellon Oregon State Michigan 2 6/26/2015 13 DOF in SS and 6 Actuators MARLO “Look ma, no bar” Brian Buss K. Akbari Hamed A. Ramezani Brent Griffin Models are Hybrid Models • Hybrid Lagrangian Dynamics SS — Single Support DS — Double Support • Underactuated Walking gait: alternating phases: SS, DS, SS, DS, … Models are Hybrid Models are Hybrid Impact Dynamics DS — Double Support Walking gait: alternating phases: SS, DS, SS, DS, … DS — Double Support Walking gait: alternating phases: SS, DS, SS, DS, … 3 6/26/2015 Models are Hybrid Models are Hybrid Walking gait: alternating phases: SS, DS, SS, DS, … SS Models are Hybrid SS DS Models are Hybrid Hybrid System: May have multiple phases “Degree” of Actuation joint angles & velocities motor torques “Degree” of Actuation joint angles & velocities motor torques No longer fully actuated! • Fully actuated: dim u = dim q Common assumption, but dangerous. • Fully actuated: dim u = dim q severe limitations on ankle torque (to prevent rotation of foot) 4 6/26/2015 “Degree” of Actuation State of the Art: DRC 6 & 7 June 2015 The bipeds in the following videos are trying to stay fully actuated! joint angles & velocities motor torques No longer fully actuated! • Fully actuated: dim u = dim q severe limitations on ankle torque (to prevent rotation of foot) ZMP, M. Vukobratovic (1968) Capture Point, Pratt et a. (2006) State of the Art: DRC 6 & 7 June 2015 State of the Art: DRC 6 & 7 June 2015 “Degree” of Actuation “Degree” of Actuation joint angles & velocities • Fully actuated: dim u = dim q • Underactuated: dim u < dim q Passive Pivot motor torques series elastic actuators • Fully actuated: dim u = dim q • Underactuated: dim u < dim q 5 6/26/2015 Periodic Orbit Poincaré map (1854-1912) Steady-state Walking Periodic Solution • How to find a periodic solution? • How to check for stability? Trajectory Tracking vs. Virtual Constraints Feedback Design • Virtual constraints • Design via parameter optimization • Initial 3D experiments Physical Virtual Constraints Trajector y Gait Generato r NL PID Cont. Virtual constraints y = (variables to be controlled) – (desired evolution) Gait phase or timing variable Feedback controller: y(t)0 6 6/26/2015 Virtual constraints Virtual constraints “holonomic constraint” Gait phase or timing variable Feedback controller: y(t)0 Virtual constraints Virtual constraints Virtual Constraints in ODE model Switching Surface Virtual constraints Virtual constraints Virtual Constraints in ODE model Virtual Constraints in ODE model Byrnes-Isidori Zero Dynamics (‘88) 7 6/26/2015 Geometric Interpretation Geometric Interpretation Hybrid Zero Dynamics In general Theorem: [TAC 2001, TAC 2009, Book] Can design surface such that Hybrid zero dynamics dim 2(N-k) Hybrid zero dynamics dim 2k • Rendering Z hybrid invariant • Rendering Z sufficiently rapidly attractive. Hybrid zero dynamics Design of the Periodic Orbit From invariance reset map is often “expansive” “Computation and Orbit Design” Adjust feedback gains to achieve attractivity of Z 8 6/26/2015 Constrained Optimization (standard) Desired Evolution y = (variables to be controlled) – (desired evolution) Design desired evolution Cost Function Integrate over Hybrid Zero Dynamics 2(N-k) dim Free parameters to be chosen Constrained Optimization (standard) Nonlinear-Hybrid Feedback Control Design desired evolution Cost Function Equality Constraints Inequality Constraints • • • • • Swing leg impact at end of step • Walking speed • Periodicity or not Ground reaction force positive Friction coefficient < 0.6 Swing foot clearance Torque bounds Stability ??? Choosing What to Control? Free Pivot • For 1 DUA, STABILITY is INDEPENDENT of CHOICE of outputs! angular vs. “Cartesian” • Otherwise (such as in 3D), choice matters. 9 6/26/2015 3D: Output Choice Matters Swing leg 3D: Output Choice Matters Stance leg Robot Falls Frontal Plane Sagittal Plane 3D: Output Choice Matters Swing leg 3D: Output Choice Matters Stance leg Eigenvalues now have magnitude less than one Robot Walks Frontal Plane Sagittal Plane 3D: Output Choice Matters Eigenvalues have magnitude greater than one Choosing What to Control Systematic Selection of Outputs Robot Falls Kaveh Akbari Hamed Brian Buss CDC 2014; ADHS 2015; IJRR [to appear] 10 6/26/2015 Systematic Virtual Constraint Selection Systematic Virtual Constraint Selection Systematic Selection of Outputs • Parameterized family of controllers Periodic Orbit • Periodic orbit: independent of parameters • Seek parameter values giving (exp.) stability Systematic Virtual Constraint Selection Orbit independent of parameter choice but changes the zero dynamics! Systematic Virtual Constraint Selection Applies Much More Broadly Periodic Orbit Feedforward term so that orbit independent of parameter choice Systematic Virtual Constraint Selection Systematic Virtual Constraint Selection 11 6/26/2015 Systematic Virtual Constraint Selection Systematic Virtual Constraint Selection Seek parameter vector so that Jacobian is Hurwitz Systematic Virtual Constraint Selection Systematic Virtual Constraint Selection Taylor Series Expansion Taylor Series Expansion Nominal parameter vector Objective: Choose such that sum of matrices is Hurwitz Choosing What to Control Seek Lyapunov function and satisfying Bilinear Matrix Inequality BMI Optimization Problem Can be handled by the solver PENBMI. 12 6/26/2015 Systematic Virtual Constraint Selection Systematic Virtual Constraint Selection Giving new outputs that couple pitch and roll in ways that we would not have found through intuition Giving new outputs that couple pitch and roll in ways that we would not have found through intuition “standard” “new” Systematic Virtual Constraint Selection Systematic Virtual Constraint Selection Robot “falls” when we cut the power. Can Include Disturbance Rejection Goal Outdoors Rough Terrain IJRR (to appear) & ADHS (submitted) 13 6/26/2015 Virtual Constraints and HZD Westervelt et al. 2004 Virtual Constraints and HZD Martin et al. Chicago Rehabilitation Institute and UT Dallas Sreenath et al. 2011 Robert D. Gregg et al., ICORR 2013 • Leg placement? No! • ZMP? No! Ames et al. 2015 Gregg et al. 2014 Buss et al. 2014 Vanderbilt leg being controlled with methods conceived at Michigan for bipedal robots. Virtual Constraints and HZD Virtual Constraints and HZD Robert Gregg Aaron Ames (Ga Tech & TAMU) Oregon State: Mikhail Jones and J. Hurst Acknowledgments University of Michigan Brian Buss, Brent Griffin, Kaveh Akbari Hamed, Ali Ramezani, Kevin Galloway, Dennis Da, Ross Hartley, Omar Harib Oregon State University Jonathan Hurst, Dynamic Robotics Laboratory, Jesse Grimes, Ross McCullough, Soo-Hyun Yoos 14 6/26/2015 Concluding Remarks Great area for feedback control There is a lot going on … … and much more to do 15
© Copyright 2026 Paperzz