Biomimetic Robots for Robust Operation in Unstructured Environments M. Cutkosky and T. Kenny Stanford University R. Full and H. Kazerooni U.C. Berkeley R. Howe Harvard University R. Shadmehr Johns Hopkins University Site visit -- Stanford University, Sept. 2, 1990 http://cdr.stanford.edu/touch/biomimetics BioMimetic Robotics MURI Berkeley-Harvard Hopkins-Stanford Main ideas: • Study insects to understand role of passive impedance (structure and control), study humans to understand adaptation and learning (Full, Howe,Shadmehr) • Use novel layered prototyping methods to create compliant biomimetic structures with embedded sensors and actuators (Cutkosky, Full, Kenny) • Develop biomimetic actuation and control schemes that exploit “preflexes” and reflexes for robust locomotion and manipulation (Full, Cutkosky, Howe, Kazerooni, Shadmehr) Issues in studying, designing and building biomimetic robots (and the basic outline for today’s site visit) 2. 1. Low-Level Control Biomimetic Robots High-Level Control MURI 3. Design & Fabrication Guiding questions What passive properties are found in Nature? Preflexes: Muscle and Exoskeleton Impedance Measurements (Berkeley Bio.) Low-Level Control High-Level Control MURI What properties in mechanical design? Biological implications for Robotics Basic Compliant Mechanisms for Locomotion (Stanford) Variable compliance joints (Harvard, Stanford) Fast runner with biomimetic trajectory (Berkeley ME) How should properties be varied for changing tasks, conditions ? Matching ideal impedance for unstructured dynamic tasks (Harvard) Fabrication Guiding questions Low-Level Control 1 cm What strategies are used in insect locomotion and what are their implications? MURI High-Level Control Fabrication Insect locomotion studies (Berkeley Bio) New measurement capabilities (Stanford) What motor control adaptation strategies do people use and how can they be applied to robots? Compliance Learning and Strategies for Unstructured Environments (Harvard & Johns Hopkins) Implications for biomimetic robots (Harvard, Stanford) dt=10ms dt=10ms dt=30ms Are preflexes enough? Guiding questions Low-Level Control MURI High-Level Control How do we build robust biomimetic structures and systems? Fabrication Shape deposition manufacturing of integrated parts, with embedded actuators and sensors (Stanford) How do we build-in tailored compliance and damping? Structures with functionally graded material properties (Stanford) Effects of Compliance in simple running machine (Stanford, Berkeley ME) Low Level Biomimetic Control 9:30-11:00 • Results on measurements of muscles, exoskeleton, compliance, damping (Full ~30) • Implications for biomimetic robots (Bailey ~20min) • Matching leg trajectory and scaling (Kazerooni ~15) • Matching impedance to dynamic task (Matsuoka ~15) What passive properties are found in Nature? Preflexes: Muscle and Exoskeleton Impedance Measurements (Berkeley Bio.) Low-Level Control High-Level Control MURI What properties in mechanical design? Biological implications for Robotics Basic Compliant Mechanisms for Locomotion (Stanford) Variable compliance joints (Harvard, Stanford) Fast runner with biomimetic trajectory (Berkeley ME) How should properties be varied for changing tasks, conditions ? Matching ideal impedance for unstructured dynamic tasks (Harvard) Fabrication MURI Year One Meeting 1999 Lower Level Control Professor Robert J. Full Daniel Dudek Dr. Kenneth Meijer University of California at Berkeley Department of Integrative Biology rjfull@socrates.berkeley.edu http://polypedal.berkeley.edu Lower Level Control Higher Centers Sensors aero- , hydro, terra-dynamic Open-loop Feedforward Controller (CPG) Mechanical System (Actuators, limbs) Feedback Controller Adaptive Controller Environment Closed-loop Sensors Behavior Chain of Reflexes Cruse Controller Inspired by Stick Insects Rough Terrain Fractal Surface Variation 3 times the height of the center of mass Control Challenge Precise Novel Slow Static Control Mechanical Gross Repetitive Rapid Dynamic Neural Feedforward Continuous Feedback (Reflexes) Feedforward Continuous Feedback (Preflexes) PolyPEDAL Control Control algorithms embedded in the form of animal itself. Control results from properties of parts and their morphology. Musculoskeletal units, leg segments and legs do computations on their own. Lower Level Control Higher Centers Sensors aero- , hydro, terra-dynamic Open-loop Feedforward Controller (CPG) Mechanical System (Actuators, limbs) Feedback Controller Adaptive Controller Environment Closed-loop Sensors Behavior Contribution to Control Mechanical System Neural System Feedforward Preflex Motor program acting through moment arms Predictive Intrinsic musculo-skeletal properties Rapid acting Passive Dynamic Self-stabilization Reflex Neural feedback loops Slow acting Active Stabilization MURI Interactions Motor Control & Learning Johns Hopkins Rapid Prototyping Stanford Muscles and Locomotion UC Berkeley MURI Manipulation Harvard Sensors / MEMS Stanford Robot & Leg Mechanisms UC Berkeley Manufactured Legs What properties should legs possess? Why? Act as springs to store and return energy? How? Act to reject disturbances? Road Map 1. System Impedance 2. Leg Impedance 3. Muscle Impedance Spring-Mass Systems EIGHT- Legged SIX- Legged Cockroach TWO-Legged Crab Body Weight Vertical Force FOUR- Legged Fore-aft Force Time Blickhan 1989 Human Dog Virtual Leg Stiffness k rel = 100 F mg TROTTERS RUNNERS HOPPERS Dx x Blickhan and Full, 1993 Human Quail 10 Dog Cockroach krel,leg 1 0.001 Hare Crab 0.01 0.1 Mass (kg) 1 Kangaroo 10 100 Sagittal Plane Model ORGANISM Spring Loaded Inverted Pendulum Multi-Leg m b k Leg Springs ? Road Map 1. System Impedance 2. Leg Impedance 3. Muscle Impedance Leg as Spring & Damper ∆x Force Stiffness, k Damping coefficient, c Restorative Forces and Perturbation Damping For an Oscillating System: Force = force due to + force due to + force due to mass stiffness damping Force = kx + . cx + .. mx Experimental Setup Oscillate Leg At Multiple Frequencies To Determine k and c Servo Motor Roach leg Length and Force recording Leg Oscillation Experiments Small Deflection at 12 Hz 0.03 0 0 -0.3 -0.03 0 0.05 0.1 Displacement Time (s) 0.15 Force 0.2 Force (N) Displacement (mm) 0.3 Leg Is Spring and Damper Small Deflection at 12 Hz 0.03 Force (N) Slope ≈ Impedance -0.3 0.3 -0.03 Displacement (mm) Effect of Frequency Impedance Increases with Frequency 0.035 Force (N) k25 Hz > k0.08 Hz -0.3 0.3 -0.035 Force @0.08 Hz Force @25 Hz Displacement (mm) Impedance Impedance of Metathoracic Limb of Cockroach Impedance (N/m) 75 70 Preferred Stride Frequency 12 Hz 65 60 55 50 45 0.01 0.1 1 10 Frequency (Hz) 100 Leg Model Standard Linear Solid k1 c • At high frequencies: Force a (k1+k2)*(displacement) k2 • At low frequencies: Force a k2*(displacement) Stride frequency (Hz) Frequency vs Speed Cockroach 20 Natural Frequency? 15 Impedance Increases 10 * Impedance Constant Alter Leg Spring Angle Take Longer Strides 5 0 0 0.2 0.4 Speed (m/sec) 0.6 Impedance Large Deflection Non-linear k 24 Hz > k0.25 Hz Perturbation Rejection Perturbation Restorative Force 4x Body Mass Discoveries 1. Insect leg behaves like a spring and damper system. 2. Strain energy is stored in the leg and returned. 3. Force – displacement relationship shows hysteresis with significant energy dissipation (50% or more). Discoveries 4. Leg impedance increases with frequency up to 12 Hz, the preferred speed of the animal. 5. Leg impedance remains constant at frequencies above 12 Hz. 6. The leg’s natural frequency is near the frequency used by the animal at its preferred speed. Discoveries 7. Insect leg could simplify control by rejecting perturbations. For a deflection of only one mm, the leg produces a force of 0.754x body mass. Road Map 1. System Impedance 2. Leg Impedance 3. Muscle Impedance MURI Interactions Motor Control & Learning Johns Hopkins Rapid Prototyping Stanford Muscles and Locomotion UC Berkeley MURI Manipulation Harvard Sensors / MEMS Stanford Robot & Leg Mechanisms UC Berkeley Manufactured Legs What properties should actuators possess? Why? Act as springs to store and return energy? How? Act to reject disturbances? Power generation? Horizontal Plane Model ORGANISM k Lateral Leg Spring b Multi-Leg m MuscleApodeme Damped Springs ? Muscle Lever Control Stimulation Servo and Force Transducer Stimulation - pattern - magnitude - phase Strain - pattern - magnitude Frequency Workloop Technique Muscle Capacity 179 Powerspace 177c Powerspace 2 Muscle Action Potentials Power 3 Muscle Action Potentials (W/kg) Stimulation phase (%) 100 60 0.0 Damper 80 Damper Spring -100.0 + 40 Motor in vivo conditions 20 Spring * in vivo conditions 0 4 6 8 10 12 14 5 Muscle Strain % 10 15 20 -200.0 Musculo-skeletal Model Preflexes Intrinsic musculo-skeletal properties Force Insect Leg Velocity Brown and Loeb, 1999 Perturbation Experiments Passive Muscle Stiffness Significant Length Increase Stimulation Servo and Force Transducer Active+Passive Force Passive Force Force increase (mN) Effect of Step Length Increase Passive resistance is significant in muscle 177c 60 Stimulated (Twitch) (n = 4) 40 20 0 0 Relaxed 1 2 Step size (%) 3 Oscillatory Perturbations Force (mN) 5 Muscle strain (%) 0.5 % Force (mN) 5 Muscle strain (%) -0.25 0 0.25 Phase angle 0 -5 Ecomplex =(DForce/Area)/strain -5 0 Time (ms) 200 Eviscous/Eelastic=tan(phase angle) Visco-elastic Properties Passive Muscle Impedance increases with frequency in muscle 179 Impedance independent of frequency in muscle 177c Significant viscous damping in both muscles. Ecomplex (N/m2) 5 x 10 tan(phase angle) 5 1 4 0.8 3 0.6 2 0.4 M179 (n=2) M177c (n=3) 1 0.2 L=1.075 0 0 50 100 Frequency (Hz) 150 0 0 50 100 Frequency (Hz) 150 Effect of Length Passive Muscle Impedance increases with length Contribution viscous damping decreases with length 10 Ecomplex (N/m2) x 10 tan(phase angle) 1 5 8 0.8 6 0.6 4 0.4 M179 (n=2) M177c (n=3) 2 0.2 f= 50 Hz 0 0.9 1 1.1 1.2 Length 1.3 1.4 0 0.9 1 1.1 1.2 Length 1.3 1.4 Perturbation experiments Impedance during workloop. Force (mN) 300 7% Locomotor pattern + Sinusoid (A=0.5%,f=200 Hz) 0 0 100 Locomotion cycle (%) 0 100 Locomotion cycle (%) Multiple Muscle System Muscle strain (%) Stimulation Phase { m177c 0 Anatomically similar muscles provide Impedance (mN) impedance m179 during different phases of the locomotion cycle! 100 Locomotion cycle (%) Discoveries 1. Passive muscle can reject perturbations. 2. Preflexes comprise passive (fixed) and active components (adjustable). 3. Passive muscle acts like a visco-elastic actuator. (Viscous damping is responsible for a significant part of total force response to perturbation.) 4. Impedance of anatomically similar muscles is distributed over the locomotion cycle. Impact on Deliverables 1. Energy storage 2. Reject perturbations 3. Simplify control 4. Penetrate new environments 5. Increase robustness Guiding questions What passive properties are found in Nature? Preflexes: Muscle and Exoskeleton Impedance Measurements (Berkeley Bio.) Low-Level Control High-Level Control MURI What properties in mechanical design? Basic Compliant Mechanisms for Locomotion Biological implications for Robotics (Stanford) Variable compliance joints (Harvard, Stanford) Fast runner with biomimetic trajectory (Berkeley ME) How should properties be varied for changing tasks, conditions ? Matching ideal impedance for unstructured dynamic tasks (Harvard) Fabrication MURI Locomotion: Biomimetic Ideology Low-Level Control • Goal: – Navigate rough terrain with simple, robust, compliant robots • Mindset shaped by Biology – – – – Tunable, passive mechanical properties Purpose-specific geometry Simple control scheme Robust components MURI Low-Level Control • Variable Compliance?: Interpreting Biological Findings Idea – Desired reaction forces depend on the environment and locomotion speed Unload Force k24 Hz > k0.25 Hz Load Displacement • How do we represent these findings? – Not traditional spring or damper elements – Energy spent per cycle independent of frequency (area enclosed by curve is the energy spent) • Results suggest hysteretic damping MURI Low-Level Control • Variable Impedance: New Design Direction What’s the difference between compliance and impedance? – Impedance refers to the relationship: dF/dx – Stiffness refers to particular impedance relationship, namely: dF/dx = k • Hysteretic Damping – Characteristic of some heterogeneous materials – Loading and unloading create different stress-strain paths – Stress-strain curve is independent of frequency • Design Implications – Compliance is mainly a function of displacement – Damping has a significant frequency dependant term MURI Low-Level Control • Variable Impedance: Design Approach Traditional Robotic Compliance – Actuator powered – Proportional feedback control - variable compliance – Complex • multiple control laws with different objectives must work together • Low bandwidth - controller delays Set + Point - k Plant Actuator Position MURI Low-Level Control • Variable Impedance: Design Approach Different Approach – Compliant member powered – Adjustable geometry - variable impedance – Simple • mechanical properties are more predictable • separate from control law • intrinsic low level stability SDM robot limb with compliance and damping • Stiffness Variable Stiffness Joint Concept Biology is telling us what mechanical properties we really need MURI Low-Level Control Sprawl 1.0: Legged Testbed • Capture the essential locomoting elements in a low DOF robot • Explore the roles of compliance and damping in locomotion • Identify areas which can be improved by SDM MURI Low-Level Control • Sprawl 1.0: Biomimetic, not just a copy Full’s research highlights certain important locomoting components – Power-producing thrust muscles – Supporting/repositioning hip joints MURI Sprawl 1.0: Thrusting Low-Level Control Cockroach Geometry Sprawl 1.0 Geometry Robotics Analysis F1 r1 0 1 0 r2 2 = F2 Femur 2 Very Low Friction Pneumatic Piston 1 Tibia Force and Workspace • • Force and Workspace Force and Workspace Full’s research on power-producing muscles 177a,c,d,e (Ahn, Meijer) Thrust production - Decoupled, compliant system MURI Sprawl 1.0: Repositioning/Supporting Low-Level Control Cockroach Geometry Sprawl 1.0 Geometry Compliant Trochanter-Femur joint g Damped, Compliant RC Servo Actuator Actuated Body-Coxa joint • • Full’s research on Trochanter-Femur joint (Dudek) Repositioning/Supporting - Decoupled, compliant system MURI Low-Level Control • Sprawl 1.0: Findings Good design and passive mechanical properties take burden off control – Compliance and damping – Simple alternating tripod locomotion scheme – Built-in posture control • Low bandwidth geometry changes – Walking, stopping, turning, and running • Need for robust components – Traditional components are not robust - poster child for SDM Sprawl 1.0: Future Work MURI Low-Level Control • Suggestions from Full – – – – • Change location of center of mass Increase gait frequency Dynamically control middle leg set points Weaken front leg force Work in Progress – Add compliant springs in parallel with constant force pistons – Replace RC servo hip actuators with more biomimetic components