Safe Control Strategies for Hopping Over Uneven Terrain Part I Brian Howley RiSE Group Meeting October 9, 2006.

Slides:



Advertisements
Similar presentations
Reactive and Potential Field Planners
Advertisements

Incremental Linear Programming Linear programming involves finding a solution to the constraints, one that maximizes the given linear function of variables.
Optimization for models of legged locomotion: Parameter estimation, gait synthesis, and experiment design Sam Burden, Shankar Sastry, and Robert Full.
Motion Planning for Point Robots CS 659 Kris Hauser.
Hybrid Systems Presented by: Arnab De Anand S. An Intuitive Introduction to Hybrid Systems Discrete program with an analog environment. What does it mean?
P. Venkataraman Mechanical Engineering P. Venkataraman Rochester Institute of Technology DETC2013 – 12269: Continuous Solution for Boundary Value Problems.
AA278A: Supplement to Lecture Notes 10. Controller Synthesis for Hybrid Systems Claire J. Tomlin Department of Aeronautics and Astronautics Department.
1. Algorithms for Inverse Reinforcement Learning 2
Uncertainty Representation. Gaussian Distribution variance Standard deviation.
1 Stochastic Event Capture Using Mobile Sensors Subject to a Quality Metric Nabhendra Bisnik, Alhussein A. Abouzeid, and Volkan Isler Rensselaer Polytechnic.
Chapter 16 Wave Motion.
Robust Hybrid and Embedded Systems Design Jerry Ding, Jeremy Gillula, Haomiao Huang, Michael Vitus, and Claire Tomlin MURI Review Meeting Frameworks and.
Nonholonomic Multibody Mobile Robots: Controllability and Motion Planning in the Presence of Obstacles (1991) Jerome Barraquand Jean-Claude Latombe.
Date:2011/06/08 吳昕澧 BOA: The Bayesian Optimization Algorithm.
MCFRoute: A Detailed Router Based on Multi- Commodity Flow Method Xiaotao Jia, Yici Cai, Qiang Zhou, Gang Chen, Zhuoyuan Li, Zuowei Li.
Prénom Nom Document Analysis: Linear Discrimination Prof. Rolf Ingold, University of Fribourg Master course, spring semester 2008.
Motion Analysis (contd.) Slides are from RPI Registration Class.
October, Scripps Institution of Oceanography An Alternative Method to Building Adjoints Julia Levin Rutgers University Andrew Bennett “Inverse Modeling.
UNC Chapel Hill S. Redon - M. C. Lin Rigid body dynamics II Solving the dynamics problems.
Chess Review October 4, 2006 Alexandria, VA Edited and presented by Hybrid Systems: Theoretical Contributions Part I Shankar Sastry UC Berkeley.
Dynamic lot sizing and tool management in automated manufacturing systems M. Selim Aktürk, Siraceddin Önen presented by Zümbül Bulut.
Planning for Humanoid Robots Presented by Irena Pashchenko CS326a, Winter 2004.
Efficient Methodologies for Reliability Based Design Optimization
Data Flow Analysis Compiler Design October 5, 2004 These slides live on the Web. I obtained them from Jeff Foster and he said that he obtained.
Development of Empirical Models From Process Data
Fast and Robust Legged Locomotion Sean Bailey Mechanical Engineering Design Division Advisor: Dr. Mark Cutkosky May 12, 2000.
CS274 Spring 01 Lecture 5 Copyright © Mark Meyer Lecture V Higher Level Motion Control CS274: Computer Animation and Simulation.
1 Collision Avoidance Systems: Computing Controllers which Prevent Collisions By Adam Cataldo Advisor: Edward Lee Committee: Shankar Sastry, Pravin Varaiya,
1cs533d-term Notes  Typo in test.rib --- fixed on the web now (PointsPolygon --> PointsPolygons)
CE 1501 CE 150 Fluid Mechanics G.A. Kallio Dept. of Mechanical Engineering, Mechatronic Engineering & Manufacturing Technology California State University,
Chapter 13 Vibrations and Waves.  When x is positive, F is negative ;  When at equilibrium (x=0), F = 0 ;  When x is negative, F is positive ; Hooke’s.
2. Solving Schrödinger’s Equation Superposition Given a few solutions of Schrödinger’s equation, we can make more of them Let  1 and  2 be two solutions.
Program Analysis Mooly Sagiv Tel Aviv University Sunday Scrieber 8 Monday Schrieber.
CS Reinforcement Learning1 Reinforcement Learning Variation on Supervised Learning Exact target outputs are not given Some variation of reward is.
Introduction to Monte Carlo Methods D.J.C. Mackay.
1 S ystems Analysis Laboratory Helsinki University of Technology Kai Virtanen, Tuomas Raivio, and Raimo P. Hämäläinen Systems Analysis Laboratory (SAL)
In Engineering --- Designing a Pneumatic Pump Introduction System characterization Model development –Models 1, 2, 3, 4, 5 & 6 Model analysis –Time domain.
1 Research on Animals and Vehicles Chapter 8 of Raibert By Rick Cory.
Advanced Programming for 3D Applications CE Bob Hobbs Staffordshire university Human Motion Lecture 3.
A Framework for Distributed Model Predictive Control
Chapter 1 Computing Tools Analytic and Algorithmic Solutions Copyright © The McGraw-Hill Companies, Inc. Permission required for reproduction or display.
Simulating Electron Dynamics in 1D
An Introduction to Programming and Algorithms. Course Objectives A basic understanding of engineering problem solving process. A basic understanding of.
Progress in identification of damping: Energy-based method with incomplete and noisy data Marco Prandina University of Liverpool.
Approximate Dynamic Programming Methods for Resource Constrained Sensor Management John W. Fisher III, Jason L. Williams and Alan S. Willsky MIT CSAIL.
Probability and Measure September 2, Nonparametric Bayesian Fundamental Problem: Estimating Distribution from a collection of Data E. ( X a distribution-valued.
Stress constrained optimization using X-FEM and Level Set Description
LEAST MEAN-SQUARE (LMS) ADAPTIVE FILTERING. Steepest Descent The update rule for SD is where or SD is a deterministic algorithm, in the sense that p and.
Hybrid Systems Controller Synthesis Examples EE291E Tomlin/Sastry.
Optimal Path Planning Using the Minimum-Time Criterion by James Bobrow Guha Jayachandran April 29, 2002.
Randomized Kinodynamics Planning Steven M. LaVelle and James J
Path Planning Based on Ant Colony Algorithm and Distributed Local Navigation for Multi-Robot Systems International Conference on Mechatronics and Automation.
Controller Synthesis For Timed Automata Authors : Eugene Asarin, Oded Maler, Amir Pnueli and Joseph Sifakis Yean-Ru Chen Embedded System Laboratory of.
Towards Adaptive Optimal Control of the Scramjet Inlet Nilesh V. Kulkarni Advisors: Prof. Minh Q. Phan Dartmouth College Prof. Robert F. Stengel Princeton.
ICCV 2007 National Laboratory of Pattern Recognition Institute of Automation Chinese Academy of Sciences Half Quadratic Analysis for Mean Shift: with Extension.
Stryker Interaction Design Workshop September 7-8, January 2006 Functional biomimesis * Compliant Sagittal Rotary Joint Active Thrusting Force *[Cham.
Safe Execution of Bipedal Walking Tasks from Biomechanical Principles Andreas Hofmann and Brian Williams.
Use or disclosure of the information contained herein is subject to specific written CIRA approval 1 PURSUIT – EVASION GAMES GAME THEORY AND ANALYSIS OF.
OSE801 Engineering System Identification Spring 2010
CS b659: Intelligent Robotics
Perturbation method, lexicographic method
Games with Chance Other Search Algorithms
Optimal control T. F. Edgar Spring 2012.
Autonomous Cyber-Physical Systems: Dynamical Systems
Sprawl Robots Biomimetic Design Analysis: Simplified Models
Alternatives for Locomotion Control
2. Solving Schrödinger’s Equation
Large Time Scale Molecular Paths Using Least Action.
Optimal Control and Reachability with Competing Inputs
Data Flow Analysis Compiler Design
Presentation transcript:

Safe Control Strategies for Hopping Over Uneven Terrain Part I Brian Howley RiSE Group Meeting October 9, 2006

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 2 Outline Motivation and Approach Vertical Hopping Control Synthesis Problem Reach Optimization Application to Vertical Hopper Hopper Ruggedness Issues –Singularity Avoidance –Control dependent phase transitions

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 3 Motivation Biomimetic principles for legged robots have been demonstrated experimentally, but quantifiable design approaches have not been established Motivating questions include: –At what speeds and over what range of surface irregularities can a robot successfully traverse? –How will changes in control, mechanical properties, or leg morphology affect performance? –Under what conditions and to what extent will additional sensors be advantageous? Goal: develop ‘first principles’ insight into motion over uneven terrain Goal: develop ‘first principles’ insight into motion over uneven and uncertain terrain

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 4 Approach Start with the simplest problem of interest: single-legged vertical hopping – Well studied for level terrain Beuhler & Koditschek prove large stable operating regime (1988) Ringrose proposes ‘self-stabilizing’ open loop control (1997) –Topic largely unexplored for variable terrain Use game theoretic approach to robot-environment interaction –Determine worst case disturbances and optimize control –Quantify performance with respect to safe operating limits Develop general analysis tools –Adapt/refine synthesis approaches for hybrid dynamical systems –Numerical tool: Maximal Invariant Safe Subset (MISS) Algorithm Extend approach to more complex problems

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 5 Why Game Theory? Two fundamental ways to deal with uncertainty: –Random process approach Requires assumptions about statistical properties (or a whole lot of data). Can be difficult to interpret if usual assumptions about normal gaussian statistics don’t apply. –Worst case analysis Provides a conservative answer Justified for low probability/high cost events: e.g. safety Game Theory provides the worst case analysis for a dynamic system by assuming continuous and discrete disturbances are applied in the worst possible way

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 6 Vertical Hopping Control hopping height – 1 st problem in running –Well studied problem on level terrain Variable terrain while hopping in place –Safety constraints  Stumbling  Joint constraints Hopping as a game –Terrain height changes within +/- limits while hopper is at apex of flight Hopper ‘wins’ by maintaining safe operation Environment ‘wins’ otherwise –Determine optimal control and disturbance strategies –Determine optimal mechanical properties

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 7 Vertical Hopping Control hopping height – 1 st problem in running –Well studied problem on level terrain Variable terrain while hopping in place –Safety constraints  Stumbling  Joint constraints Hopping as a game –Terrain height changes within +/- limits while hopper is at apex of flight Hopper ‘wins’ by maintaining safe operation Environment ‘wins’ otherwise –Determine optimal control and disturbance strategies –Determine optimal mechanical properties

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 8 Vertical Hopping Control hopping height – 1 st problem in running –Well studied problem on level terrain Variable terrain while hopping in place –Safety constraints  Stumbling  Joint constraints Hopping as a game –Terrain height changes within +/- limits while hopper is at apex of flight Hopper ‘wins’ by maintaining safe operation Environment ‘wins’ otherwise –Determine optimal control and disturbance strategies –Determine optimal mechanical properties

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 9 Vertical Hopping Control hopping height – 1 st problem in running –Well studied problem on level terrain Variable terrain while hopping in place –Safety constraints  Stumbling  Joint constraints Hopping as a game –Terrain height changes within +/- limits while hopper is at apex of flight Hopper ‘wins’ by maintaining safe operation Environment ‘wins’ otherwise –Determine optimal control and disturbance strategies –Determine optimal mechanical properties ?

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 10 Hopper Dynamics Hybrid system –Discrete events change dynamics Flight phase –Ascent, Step, Descent phases accommodate changes in terrain height Contact phase –Thrusting may occur any time in contact –Liftoff transition (not shown) if ground reaction forces less than zero

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 11 Safety The safe set, F, is the subset of the discrete and continuous state space considered to be safe: The unsafe set, G, is the complement of F: Vertical Position Vertical Velocity Safe States, F Unsafe States, G (minimum leg length constraint) Vertical Velocity Safe States, F Vertical Position Unsafe States, G (‘Stumble’ Condition)

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 12 Control Synthesis Problem Given a safe set, F, determine (i) the maximal controlled invariant set contained in F, and (ii) the controller which renders this set invariant General Algorithm: Start with the safe set, F, and iterate backwards in time eliminating those states that can be rendered unsafe Reference: Tomlin, Claire, and Lygeros, John, and Sastry, Shankar, “A Game Theoretic Approach to Controller Design for Hybrid Systems”, Proceedings of the IEEE, Vol. 88, No. 7, July 2000.

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 13 Reach Operation Reach(G,E): the set of states that can be driven to the set of unsafe states, G, prior to reaching a (safe) escape set, E. –The Reach operation is formulated like a pair of pursuit-evasion games System dynamics: Two objective functions: q=‘Contact’ Vertical Velocity Vertical Position G  G (x)<0 E  E (x)< ( x >0, non ‘contact’ region)

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 14 Optimization-Reach Use optimal control theory to determine ‘least restrictive’ control and worst case disturbance inputs. –Hamilton-Jacobi equations: Since objective functions, J x, depends only on the terminal state, x(t f ), the optimal control and disturbance inputs, u*, d*, are such that the gradient of the objective function is perpendicular f q (x,u*,d*).  Don’t have to integrate the Hamilton-Jacobi equations!!

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 15 Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E J G Optimization at t=t f G, J G =  G q=‘Contact’

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 16 Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ J G Optimization at t=t f G, J G =  G Gradient non-zero and perpendicular to boundary

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 17 Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ Extrema point, x* J G Optimization at t=t f G, J G =  G Gradient non-zero and perpendicular to boundary Find extrema points, x*:

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 18 Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ J G Optimization at t=t f G, J G =  G Gradient non-zero and perpendicular to boundary Find extrema points, x* Backward propagation Find u*, d* at each time step u min u max Initial Condition Boundary

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 19 Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ J G Optimization at t=t f G, J G =  G Gradient non-zero and perpendicular to boundary Find extrema points, x* Backward propagation Find u*, d* at each time step u min u max Initial Condition Boundary

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 20 J G Optimization at t=t f G, J G =  G Gradient non-zero and perpendicular to boundary Find extrema points, x* Backward propagation Find u*, d* at each time step Forward propagation Find u*, d* at each time step  Optimization reverses sign Optimization-Reach Vertical Position Vertical Velocity G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ u max u min Exit Condition Boundary

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 21 J G Optimization at t=t f G, J G =  G Gradient non-zero and perpendicular to boundary Find extrema points, x* Backward propagation Find u*, d* at each time step Forward propagation Find u*, d* at each time step  Optimization reverses sign Optimization-Reach Vertical Position Vertical Velocity G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ Exit Condition Boundary

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 22 J E Optimization E is subset of Exit Conditions ( EC ) boundary Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ Exit Condition (EC) boundary

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 23 J E Optimization E is subset of Exit Conditions ( EC ) boundary At t=t f E, J E =  E, x  EC Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 24 J E Optimization E is subset of Exit Conditions ( EC ) boundary At t=t f E, J E =  E, x  EC Gradient non-zero at border Extrema points, x*, at border Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ Extrema points, x*

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 25 J E Optimization E is subset of Exit Conditions ( EC ) boundary At t=t f E, J E =  E, x  EC Gradient non-zero at border Extrema points, x*, at border Backward propagation Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ u min u max u min u max u min u max

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 26 J E Optimization E is subset of Exit Conditions ( EC ) boundary At t=t f E, J E =  E, x  EC Gradient non-zero at border Extrema points, x*, at border Backward propagation Find u*, d* at each time step Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 27 Combined Optimization Find intersection points Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 28 Combined Optimization Find intersection points Intersect state spaces Optimization-Reach Vertical Position Vertical Velocity E G Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 29 Combined Optimization Find intersection points Intersect state spaces Optimization-Reach Vertical Position Vertical Velocity E Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ Reach(G,E) Safe Subset Reach(G,E) The boundaries of the safe subset constitute the least restrictive control. Within the subset any control action u min <u<u max is safe until the state reaches the boundary. u min u max

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 30 Combined Optimization Find intersection points Intersect state spaces Backwards Chaining The new initial conditions modify the escape set E for predecessor modes Optimization-Reach Vertical Position Vertical Velocity E Consider J G and J E optimization separately and combine results: let t f G < t f E q=‘Contact’ Reach(G,E) Safe Subset u min u max New set of safe Initial Conditions

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 31 Application to Vertical Hopper Contact Phase Descent Phase Step Phase Ascent Phase Hopper Phase Plane Position vs Velocity

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 32 Application to Vertical Hopper Contact Phase Extrema point where f(x,u,d) is tangent to constraint surface Joint Compression Constraint at -0.5

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 33 Application to Vertical Hopper Backward time propagation at u = u max To Contact phase Initial Condition Forward time propagation at u = u max To Contact phase Exit Condition Contact Phase Propagation

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 34 Application to Vertical Hopper Descent Phase backward propagation from Initial to Exit conditions Descent Phase backward propagation from Exit to Initial conditions Descent Phase Propagation

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 35 Application to Vertical Hopper Shift for Step Up in backward time Shift for Step Down in backward time Step Phase

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 36 Application to Vertical Hopper Ascent Phase Propagation

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 37 Application to Vertical Hopper Contact Phase Propagation u=u min u=u max Descent Phase Propagation

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 38 Application to Vertical Hopper 2 nd Iteration through Contact and Descent Phases

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 39 Application to Vertical Hopper 3rd Iteration through Contact and Descent Phases

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 40 Application to Vertical Hopper 4th Iteration through Contact and Descent Phases

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 41 Application to Vertical Hopper Final Boundary for Safe States u=u min u=u max u=u min Safe Subset Boundaries of The Safe Subset Determine the Least Restrictive Control

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 42 Vertical Hopper Results Safe Subset Safe Subset Environment can force crash or stumble Environment can force stumble Low damping Moderate damping

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 43 Ruggedness Ruggedness is a measure of terrain variability the hopper can tolerate safely. For moderate to high damping ruggedness is maximized with soft leg springs. These results support biomimetic design principles: Control should be divided between both active and passive elements

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 44 Issues Singularites –At singular points ( f(x,u,d)=0 ) backward time propagation is undefined –General approach is to define a boundary or new phase that prescribes u, d so that singularities are avoided The implementation will be domain specific and I don’t know how to “program” a general solution What if settling to a singular point is “happy”? Control/Disturbance dependent phase transitions –Example: transition from sliding to non-sliding contact depends on control force, then changing values of control induces thrashing between modes –General approach is to define intermediate phases which fix control inputs to avoid thrashing This same restrictions probably can’t be imposed on disturbance inputs.

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 45 Singularity Avoidance Singular points for 0 ≤ u ≤ u max Approach: If f(x*,u*) trajectory reaches singularity avoidance boundary then set u=u rtn (magenta) or u=0 (blue). Control is least restrictive in the sense that initial/exit subboundaries are unchanged. Alternative: Could set u=u max to ‘blow through’ any singularities

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 46 Control Dependent Transitions Transition from contact to ascent when ground reaction force is zero: –GRF = ky+by’+u Define a liftoff phase where u=u max in transition from contact to ascent and u=0 in transition from ascent to contact. Forward in time from A1, Propagation is bound by Points at A2 and A2’. Backward in time from B1, propagation is bound By points at B2 and B2’

10/09/06 BJH Safe Control Strategies for Hopping over Uneven Terrain 47 Conclusions At least for simple systems, a game theoretic approach can be applied to gain useful insight to system limitations and behaviors in environment. –More complicated systems require more complicated analysis tools (topic for part II). –Theoretically, the approach generates a control law but implementation issues (delays, etc.) have not been addressed here. The approach has difficulty with singularities and control/disturbance dependent phase transitions which can be addressed with some thought and some restrictions.