.. in the human environment · real-time, many degrees of freedom Compatibility with Human...
Transcript of .. in the human environment · real-time, many degrees of freedom Compatibility with Human...
1
Human-Centered Robotics
Oussama KhatibArtificial Intelligence Laboratory
Department of Computer ScienceStanford University
.. in the human environment
Exploring the Human Environments
Safety & Performance
Sensing and Perception
Planning and Controlreal-time, unstructured world
real-time, many degrees of freedom
Compatibility with Human
Mechanisms and ActuationHuman-like skills
Interactivity & Human-Friendly
Safety
2
Safety
Performance
Competing?
Requirements
Human-Friendly Robots Human-Friendly Robots
• Dependable & Safe• Soft Actuators• Light Structures• Impact-reduction Skin• Low Reflected Inertia• Distributed Sensing• Good Performance
PumaConventional Geared Drive:
• Lighter structure• Large reflected
actuator inertia
JmotorNgear
Jlink
JmotorNgear
Jlink
(Jlink + N2Jmotor)Effective Inertia
Heavy structure
Technology Why Are Robotic Arms Unsafe?
Robot Collision
Head Injury Criteria (HIC)
Head Injury Criteria (HIC)
Inte
rface
Stif
fnes
s (k
N/m
)
Head Injury C
riteria Index
Effective Inertia (kg)
Puma
>20 cm compliant covering
Torq
ueM
agni
tude
Frequency
Actual Torque Requirements
Actuation Requirements
Torque Vs Frequency: Square Wave
+
Assumed Torque Requirements
Torq
ueM
agni
tud e
Frequency
3
ElasticCouplingElastic
Coupling Large BaseActuator
Large BaseActuator
Parallel Actuation
Small JointActuator
Small JointActuator
Distributed Macro Mini (DM2) Approach Robot Characteristics
Equivalent Mass-Spring Model
Robot Collision
• Effective Inertia• Effective Stiffness• Impact Velocity
Manipulator Safety Index (MSI) DM2 New Testbed
“the high capacity of a large robot with the fast dynamics and safety of a small one”
DM2 vs. PUMA 560: Effective Mass Comparison
PUMA 560 (link 2 and link 3)
HFR (conventional actuation)
HFR (DM2 actuation)
Maximum effective mass:
PUMA 560: 24.37 kg
HFR (Macro actuation): 12.71 kg
HFR (DM2 actuation): 2.81 kg
10
20
30
30
210
60
240
90
270
120
300
150
330
180 0
Manipulator Safety Index (MSI)
DMM: 10x reduction in effective inertia
4
10x reduction in effective inertia3x increase in position control bandwidth10x decrease in trajectory tracking error
SafetyAND
Performance
DistributedMacro-MiniActuation
DM2
DM2 Performance
• Whole-body control strategies
• Multi-contact and constraints
• Walking and Manipulation
Human-Like Structuresbranching and under-actuated
Task & Posture Decomposition
Whole-body Control
Task Dynamics and Control
Task Control
T FJΓ =
Task Dynamics
)ˆ ˆ ˆ( TaskV pF μ= Λ + +−∇ TaskVF −∇=F
x
xΛ F=pμ+ +
Task Dynamics – Branching Structures
x J⇒
1
2
.
m
xx
x
x
⎛ ⎞⎜ ⎟⎜ ⎟=⎜ ⎟⎜ ⎟⎜ ⎟⎝ ⎠ ⇒ Λ
5x
5f
2x2f
3x
3f1x 1f
6x6fxΛ F=pμ+ +
12 1
21 2
11
22
1 2
...
......................
...
L
L
L L LL
Λ Λ⎛ ⎞⎜ ⎟Λ Λ⎜ ⎟Λ =⎜ ⎟⎜ ⎟Λ Λ⎝ ⎠
ΛΛ
Λ
5
Control Structure
task postuT Ttask task reJ NF= +Γ Γ
0taskx =⇒
Decomposition of torque vector
Dynamic Consistency
Task and Posture Control
Whole-body Control
Task Energy
Posture Energy
no joint trajectories⇒
DynamicallyDecoupled
Learning from the human
Posture Energy?
Human Natural Motion
Motion Capture Specific MotionMotion Characteristics
HumanMotionCapture
HumanDynamicModel
Musclephysiology
Skeletalphysiology
Marker dataHuman motion
Human Motion Characterization
6
Motion capture
Simulation 79 DOF and 136 MusclesBiometric Data & Bone Geometry
Dynamic simulation
In motion, humansminimize fatigue, under physical and “social” constraints
Learning from the Human
Physiology-based Posture Field
Physiology-based Posture Field
T FJΓ =TL mΓ =
A Task, F:
Muscle actuation:
cNMuscle capacities: Configuration-dependenttorque bounds
Physiology-based Posture Field
Human posture is adjusted to reduce muscular effort
2E cm=Human-muscular Energy minimized:
Function of physiology, mechanical advantage, and task
2 1)( ) ]([Tc
TTL NE q F J L FJ− −=
Data from Subjects
8
SAI EnvironmentDynamic simulation, control, & haptics
SAI Neuromuscular Library
Constraints and Priorities
Self Collision
Collision Avoidance
Collision Avoidance
9
Elastic Planning Elastic Planning
Interactive Haptic Simulation Contact/Collision Resolution
Multi-point contactmulti-articulated body
Single point contactwith a single body
Interactive
Efficient Dynamic AlgorithmsSimulation, Contact Resolution, and Control
m1 m2
10
Crash TestsIntegrationof Walking
Multi-Contact Whole-body ControlIntegration of Whole-Body Control & Walking
Under-actuated Balance Reaction forces
Multi-Contact Whole-body ControlIntegration of Whole-Body Control & Walking