Therapeutic Robotics

Neville Hogan

Sun Jae Professor of Mechanical EngineeringProfessor of Brain and Cognitive Sciences

Massachusetts Institute of Technology

Disclosure: Neville Hogan is part owner of Interactive Motion Technologies, Inc. which manufactures human-interactive technologies under license to MIT.

Newman Laboratory for Biomechanics and Human Rehabilitation

Contact Robotics

Robot apprentices/helpers

Robot big brothers/bodyguards

Robot dance partners

Robot nurses

Robot therapists

...and many more

My enduring interest: controlling physical interaction

Physical contact can make control extraordinarily difficult……but affords powerful new task strategies

Ultimate goal: Close physical cooperation

between machines & humans

Image courtesy of Prof. Kazuhiro Kosuge

Robotic Therapy

More than twice the benefit of conventional therapy alone

Fewer side effects (joint pain)

Contact and interaction are essential

Meta-Analysis Demonstrates Reliable Benefits

Kwakkel et al. (2008) “Effects of Robot-assisted therapy on upper limb recovery after stroke: A Systematic Review,”

Neurorehabilitation and Neural Repair

22(2):111-121.

Most Effective Therapy Requires Variable Impedance

Controller sets up a “virtual slot”

between start and target

Springy slot walls provide permissive guidance

Deters inappropriate movement (assists aiming)

Assist only as needed

Back wall closes in on target

Allows free movement towards target

Adaptive impedance control

Track patient’s performance

Adjust wall stiffness and allotted move time

Continually challenge the patient

Krebs et al. (2003) Rehabilitation robotics: performance-based progressive robot-assisted therapy. Autonomous Robots

15:7-20.

Contact Robotics requires “High-Force”

Haptics

Present Actuator Technologies

Electromagnetic: low force density

Hydraulic, geared electromagnetic: high intrinsic impedance

Compressed-gas: limited by low-

frequency “parasitic”

dynamics

Feather-light touch at forces up to and beyond body weight

High force density (force/mass ratio)

Low output (driving-point) mechanical impedance

The Appeal of Force Feedback

Equation of motion:

Force feedback controller:

Resulting equation of motion:

Increasing Gf

reduces apparent inertia, friction

eaf FFxxFxm ),(

efa FGF

ef

f

f

FG

xxFGxm

1

),(1

SNAG—Coupled Instability

Performance Comparison

Best-Practice Prior Art

New Optimized Controller

Buerger, S.P. and Hogan, N. Complementary Stability and Loop-Shaping for Improved Human-Robot Interaction. IEEE Transactions on Robotics 2007 23:232-244

Goal: feather-light touch at high force

Controller instability has plagued prior approaches

Progress: new force-control strategy

Low impedance without instability

US & foreign patent applications 2006

Achieved PerformancePerformance best tested by “feel”

Static/Coulomb friction:less than 0.3 N (66x reduction)

Inertia reduction:1.75 kg with Kdc

=2000 (3.4x reduction)

1.2 kg with Kdc=3000 (5x reduction)

Performance and stability are significantly enhanced

—despite differences between model and robot.

Robotic Lifting Therapy

45 year-old right-

handed male, suffered stroke on 5/2000

Left subcortical infarct involving basal ganglia & internal capsule

Complete right hemiplegia before robotic therapy

Robotic Locomotion Therapy: Mobile Overground Prototype

Mobile robot follows patient

Overground or can “park”

at a treadmill

Platform for therapy robot modules

Body weight support (continuously variable)

Interact with pelvis to challenge and/or treat balance, weight shifting …

... and more

Main design challenge:

Variable impedance actuators

Compressible Fluids

Piston & cylinder with ideal gas, isothermal conditions

Adiabatic similar

Force varies inversely with x

Stiffness varies inversely with x2

But…

Stiffness is proportional to mass of gas, m

Abrupt stiffness change requires excessive mass flow rate

e.g. virtual wall contact is limited by choked flow

Electro-Hydraulics

High working pressure

excellent force/mass ratio

Flapper valves

Roa too small insufficient control Pa ≈

Ps

Rlb increases as Rla decreases

Low output Z requires both Rlband Rla low excessive leakage

High intrinsic output mechanical impedance

Jet-pipe valves similar

Hybrid Hydraulic Actuator

A hybrid (compromise) solution

Electro-mechanical motor

Mounted on robot base where weight is less problematic, even advantageous

Flexible hydraulic fluid line

Piston-cylinder actuators

Low end-effector force/mass ratio

Hybrid Design Enables Impedance Shaping

Integrate dynamics with fluid transmission

Spring & damper via flexible membrane

Variable inertia also plausible

Unrestrained surface determines impedance magnitude

Variable impedance by varying restraint

Sliding collars restrict membrane deformation

Dynamically “clean”

implementation

Progress so far…

Improved endpoint force/weight ratio

>3x better than linear motor

Acceptable stiffness range:

0 to 3,700 N/m

Acceptable force bandwidth:

~20Hz, n

≈8Hz ≈0.1-0.2

Endpoint inertia and static friction could be improved

Newman Laboratory for Biomechanics and Human Rehabilitation

Ankle robotAnti-gravity robot

Shoulder & elbow robot

Wrist robot

Whole-arm robot

Hand robot

Therapeutic Robotsfor

Neuro-Recovery

CollaboratorsHermano KrebsBruce VolpeMindy AisenFletcher McDowellJoel SteinWalter FronteraAlbert LoChristopher BeverRichard MackoLarry ForresterMargaret FinleySusan FasoliLaura diPietroBrandon RohrerAnindo Roy

SupportNational Institutes of HealthNational Science FoundationNational Institute for Disability and

Rehabilitation ResearchVeteran’s AdministrationNational Defense Sciences Education GrantSamsung FellowshipBurke Medical Research InstituteLangeloth FoundationGloria Blake Memorial FundEric P. & Evelyn E. Newman Fund

Jerome PalazzoloStephen BuergerDustin WilliamsJames CelestinoKristin JugenheimerLisa EdelsteinMark FerraroRichard HughesChrista DielsJennifer KrolDaniel LynchRita PopatMike RobertsJason WheelerLorenzo Masia

Newman Laboratory for Biomechanics and Human Rehabilitation

Acknowledgements

Coupled Stability via Passivity

A passive

impedance has Z(s) positive real

Phase of Z(s) lies between +90°

and -90°

System may store, dissipate & return energy—

but cannot be “pumped”

to supply power continuously.

Physical interaction resembles unity negative feedback

Couple two passive systems

Combined phase lies between +180°

and -180° STABLE

Controller design constraint:

Imposing

passive robot impedance

guarantees stability when coupled to all

passive objects.

Arbitrary complicated collections of springs, masses, dampers, constraints, etc.

Colgate, J. E. and Hogan, N. (1988) Robust Control of Dynamically Interacting Systems, International Journal of Control, Vol. 48, No. 1, pp. 65-88.

Hogan, N. (1988) On the Stability of Manipulators Performing Contact Tasks,

IEEE Journal of Robotics and Automation, 4: 677-686.

Force Feedback and Passivity

Passivity is hard to achieve

Discrete-time implementation exceeds phase constraint at high frequencies

High-gain force feedback with resonant dynamics between sensor & actuator violates passivity

Passivity is conservative

With any

resonant dynamics between sensor & actuator, force feedback inertia reduction by 50% or more is non-

passive [Colgate ’89]

Severely

limits force feedback loop gain

Complementary Stability

Define a bounded set of environment port functions:

Definition: A robot represented by Z

achieves complementary stability with the set Y

if the

coupled system is robustly stable

Stability analysis by the small gain theorem

Additive perturbation structure is not essential

)()()()( ssWsYsY n 1)(

s

Controller Design via Constrained Optimization

Prerequisites

Model of robot (with at least one resonance)

Model (or data representation) of environment port admittance

Assumed controller structure with selected variable parameters

Algorithm

Broad search finds parameter combinations to satisfy complementary stability

Select best-performing stable controller(s) based on robot impedance magnitude

Controller Design Example

Robot model:

Single-resonance, with force transducer

Control structure:

Vary p, z, Kdc

Target impedance Z=0

Environment model:

Stability by structured singular value

Performance “cost”:

Parameters based on laboratory robot module, literature on human arm endpoint dynamics

edca Fpszs

zpKF

)()(

1

0

)(log

jZC

Example ResultsCost C at maximum stable Kdc

versus p

and z

Region “A”

indicates low-frequency lag control

Region “B”

indicates high-frequency lead control

Example Results (continued)

Algorithm returns non-obvious

controller parameters.

Implementation

Apply control to physical system

Screw-driven robot module

140 N continuous force capacity

Up to 20 N Coulomb friction, position dependent

Approximately 6 kg endpoint inertia

High-frequency noise in force sensor precludes high-frequency (lead) control

Model is linear, robot significantly not

Robust testbed for control approach

Stability

Contact tests with spring (and plastic block) environments

Last column indicates behavior coupled to human arm

“-”

indicates unwanted vibration

Model-based algorithm results are more conservative than experiments, less conservative than passivity

Instrument Panel InstallationInstrument Panel InstallationEssential Partner Robot TechnologyEssential Partner Robot Technology

Skill AssistSkill Assist

・Little Force Required・Little Force Required・Simple, Precise Positioning・Simple, Precise Positioning

Reduce Manpower for Heavy Lifting TasksReduce Manpower for Heavy Lifting Tasks