Biomimetic Robots for Robust Operation in Unstructured Environments

1
Biomimetic Robots for Robust Operation in
Unstructured Environments
BioMimetic Robotics MURI Berkeley-Harvard Hopkins
-Stanford

• R. HoweHarvard University
• M. Cutkosky and T. KennyStanford University
• R. Full and H. KazerooniU.C. Berkeley
• R. Shadmehr
• Johns Hopkins University

http//cdr.stanford.edu/touch/biomimetics
ONR/DARPA MEETING ON LEGGED ROBOTS, COOPERATIVE
BEHAVIOR, AND NAVIGATIONCOSTAL SYSTMS STATION,
PANAMA CITY, MAY 4-5, 1999
2
Main ideas
BioMimetic Robotics MURI Berkeley-Harvard Hopkins
-Stanford

• 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)

3
Status - Locomotion
BioMimetic Robotics MURI Berkeley-Harvard Hopkins
-Stanford

• First year of project
• Preliminary experiments to determine insect leg
  properties
• Fabricated first prototypes of embedded sensors
  and actuators
• Locomotion focus rough terrain traversal,
  inspired by cockroach running over blocks
  surface 3x shoulder height

4
MURI Interactions Areas and Leadership
Rapid Prototyping
Stanford
Motor Control
Muscles and
Learning
Locomotion UC Berkeley Bob Full
Johns Hopkins
MURI
Robots Legs
Manipulation
UC Berkeley
Harvard
Sensors / MEMS
Stanford
5
Neuro-Mechanical Model
Aero-, hydro-, terra-dynamic
Higher
Sensors
Environment
Centers
Open-loop
Behavior
Mechanical
Feedforward
System
Controller
(CPG)
(Actuators, limbs)
Feedback
Closed-loop
Controller
Adaptive
Sensors
Controller
6
Neuro-Mechanical Model
Feedforward
Mechanical
Behavior
Controller
System
(CPG)
(Muscles, limbs)
Closed-loop
Sensors
Reflexive Neural Feedback
7
Contribution to Control
Neural System
Mechanical System
Reflex
Feedforward
Preflex
Motor program acting through moment arms
Intrinsic musculo- skeletal properties
Neural feedback loops
Slow acting
Predictive
Rapid acting
Passive Dynamic Self-stabilization
Active Stabilization
8
Perturbation Response
Working Hypotheses
Force Perturbation Animal Strikes Obstacle
Smaller Reaction Force Joint Angles Altered Less
Stable Decreased Speed
No Preflex
Preflex
Larger Reaction Force Joint Angles Similar More
Stable Maintain Speed
9
Discoveries

• Preflex Present
• No Active Reflex Required
• Stiffness Varies During Cycle

10
Perturbation Experiments
Muscle is Stiffest at Midstance
11
Leg Stiffness
1st Measures of Leg Stiffness, Damping
Servo Motor
Roach leg
Length and Force recording
12
Impact on Deliverables
1. Flexible Robot Leg Could Reject
Perturbations 2. Simplify Control
(feedforward) 3. Suggest Design of Artificial
Muscles
13
Micromachined Force Sensor for Adhesion Force
Measurement of Single Gecko Setae
MEMS Instrumentation for biomechanics studies
(Kenny/Full)

• Yiching Liang and Tom Kenny
• Stanford University

106 setae per animal, average 4.7 ?m diameter
Wall climbing mechanisms Suction, Capillary
(wet) adhesion, Micro-interlocking,
Electrostatic attraction - NOT van der Waals
forces?
14
2-Axis Micromachined Force Sensor
Special 45? ion implantation to embed
piezoresistors on surfaces and side walls.
Attachment point
Gecko measurements now underway...
15
MURI Interactions Areas and Leadership
Rapid Prototyping
Stanford
Motor Control
Muscles and
Learning
Locomotion UC Berkeley
Johns Hopkins Reza Shadmehr
MURI
Robots Legs
Manipulation
UC Berkeley
Harvard
Sensors / MEMS
Stanford
16
Neuro-Mechanical Model
aero- , hydro, terra-dynamic
Higher
Sensors
Environment
Centers
Open-loop
Behavior
Mechanical
Feedforward
System
Controller
(CPG)
(Actuators, limbs)
Feedback
Closed-loop
Controller
Adaptive
Sensors
Controller
17
Relating Limb Impedance and Learning
General Goal Understand human arm impedance
strategies when learning tasks in unstructured
environments Challenges The biomechanics of the
human arm are dominated by multiple time delays
in feedback. How do time delays affect
measures of arm impedance? Humans learn internal
models to learn control. How does a change in
the internal model affect measures of arm
impedance?
18
Results
In general, time delays in feedback reduce
apparent viscosity and add apparent mass to a
system. Example
19
Human Arm Motor Control Model
A model of the human arms time-delayed control
processes were used to derive bounds on the
impedance changes that should occur as a function
of learning.
20
Implications for Robot Control

• Relates delays to variation in limb impedance -
  convenient means of analyzing mechanical
  interactions
• Method for trading off costs of higher-level
  processing delay vs. passive impedance

21
MURI Interactions Areas and Leadership
Rapid Prototyping
Stanford
Motor Control
Muscles and
Learning
Locomotion UC Berkeley
Johns Hopkins
MURI
Robots Legs
Manipulation
Harvard Robert Howe
UC Berkeley
Sensors / MEMS
Stanford
22
Impedance in Manipulation

• Example Grasping in an unstructured environment
• Before contact No interaction force gt Low
  arm stiffness k
• Collision with object produces small disturbance
  force

Muscle Impedance
f k x
23
Variable Impedance Manipulation Testbed

• Whole-Arm Manipulator (Barrett Technologies)
• Low moving mass
• Negligible friction
• Back driveable
• gt Low impedance robot

24
Goal Minimum Impedance Grasping and Maniplation

• Combine biologically-inspired elements
• low-impedance manipulator
• feedforward dynamic models (limit feedback
  gains to reduce impedance)
• simple contact sensing

Intrinsic tactile sensing(contact location
from force-torque measurements)
gt Ability to probe and grasp objects with
minimum forces in unstructured environments
25
MURI Interactions Areas and Leadership
Rapid Prototyping
Stanford
Motor Control
Muscles and
Learning
Locomotion UC Berkeley
Johns Hopkins
MURI
Robots Legs
Manipulation
UC Berkeley Hami Kazerooni
Harvard
Sensors / MEMS
Stanford
26
Objectives

• Create a robust, simple, and fast legged
  platform, able to traverse rough block surface
• Use off-the-shelf fabrication technology
• Explore role of open-loop impedance and
  mechanical design
• Serve as early testbed for control concepts

27
Initial Focus Leg Mechanism
Full has shown that a substantial portion of
locomotor control is simple and resides in the
mechanical design of the system

• Biological Observations
• Control results from the properties of the parts
  and their morphological arrangement.
  Musculoskeletal units and legs do much of the
  computations on their own by using segment mass,
  length, inertia, elasticity, and damping as
  primitives.

• Engineering Equivalence
• System performance is function of the physical
  system no feedback control has been used to
  alter the dynamics of the system.

28

• Biological Observations
• Position control using reflexes is improbable if
  not impossible
• During climbing, turning, and maneuvering over
  irregular terrain, animals use virtually the same
  gait as in horizontal locomotion - an alternating
  tripod. The animals appear to be playing the same
  feedforward program for running.

• Engineering Equivalence
• No need for sensors for position speed, or force
  control
• A one degree of freedom system only. No need to
  design elaborate multi-variable robotic legs.

29
(No Transcript)
30
1-DOF Linkage Design Example
f
g
b
c
a
d
31
(No Transcript)
32
MURI Interactions Areas and Leadership
Rapid Prototyping Stanford Mark Cutkosky
Motor Control
Muscles and
Learning
Locomotion UC Berkeley
Johns Hopkins
MURI
Robots Legs
Manipulation
UC Berkeley
Harvard
Sensors / MEMS Stanford Tom Kenny
33
Application Small robots with embedded sensors
and actuators
Building small robot legs with pre-fabricated
components is difficult Is there a better way?
34
Shape Deposition Manufacturing (CMU/SU)
Deposit (part)
Shape
Shape

• Embedded Components Soft materials gt
• Improved robustness
• Simplified construction

Deposit (support)
Embed
35
Robot leg example(http//cdr.stanford.edu/biomime
tics)
Steel leaf spring
Designer composes the design from library of
primitives, including embedded components
Piston
Part Primitive
Outlet for valve
Valve Primitive
Circuit Primitive
Inlet port primitive
36
Robot Leg compacts
The output of the software is a sequence of 3D
shapes and toolpaths.
Embedded components
Part
Support
37
Robot Leg embedded parts
Steel leaf-spring
Piston
Sensor and circuit
Valves
A snapshot just after valves and pistons were
inserted.
38
Pressure Control in Small Pneumatic Systems

• SDM allows fabrication of small integrated
  mechanisms
• Control of small pneumatic systems with
  off-the-shelf components (solenoid valves) is in
  a challenging regime
• Miniature analog servo-valves needed for smooth
  performance are not available

t
line
t
Solenoid Valves
Pressure Control Impossible
Equal
Performance
PWM Control
t
volume
Small Pneumatic Systems
Usual regime of Operation
39
Different Sensors and Actuators have different
considerations for embedding, generally these
include
SDM Considerations for Embedded Sensors/Actuators

• Coupling and Adhesion
• Fixturing, Positioning, Placement
• Protection and Encapsulation
• Multiplexing, Connectivity, Interconnect
  Integrity and Strain Relief
• Thermal energy generation and cooling

40
Sensor circuit boards - interconnect pins
protected in wax before embedding Circuit
boards embedded with pressure sensor--sensor
ports protected with wax
41
Embedded sensor and circuitry with sacrificial
wax removed Assembled into pneumatic system
42
Robot Leg completed
Finished parts ready for testing
43
MURI Interactions Areas and Leadership
Rapid Prototyping
Stanford
Motor Control
Muscles and
Learning
Locomotion UC Berkeley
Johns Hopkins
MURI
Robots Legs
Manipulation
UC Berkeley
Harvard
Sensors / MEMS
Stanford