Adaptive Locomotion Control: From Animals to Robots
CH3: Biomechanics and Neural Mechanisms of Robots
Poramate Manoonpong ([email protected])www.manoonpong.com/ALCLecture/CH3
• Poramate Manoonpong (POMA, [email protected] )
• Shao Donghao ([email protected] )
• Zhang Yanbin ([email protected] )
• Sun Tao ([email protected] )
• Dong Yi ([email protected] )
• Potiwat Ngamkajornwiwat ([email protected] )
• Worasuchad Haomachai ([email protected] )
• Kobe Kang ([email protected] )
• 3 credits
• Autumn 2020 (4 hours / Block )→ One Block/ week (8:00 to 12:00)
• Lectures, exercises, a project for your report
Lecture (Theory: up to 2-3 hours of each block):
Exercises & project (Practice: Robot simulation):
Link: https://mooc1.chaoxing.com/course/215088272.html?edit=true&articleId=2274940122
About the course
3
• Evaluation: 20% (exercise, attendance, Q&A) + 80% Final exam
• Final Exam: individual written report based on project and evaluated with an external censor. Your report must not be identical to your colleague!!
• Assessment: report (as a paper) presents “locomotion control mechanisms and result”(deadline by January 5th, 2021) , the report must be submitted as a pdf file via email to [email protected]
→ Please wrtie your email subject as ”Submit ALC report 2020”
• The report (11-page MAX length) with the given format must contain:
1) Abstract, 2) Introduction, 3) Materials and Methods, 3) Experiments and Results (link to your results, e.g., on web), 4) Discussion & Conclusions, 5) Acknowledgements, 6) References
>> You can get feedback from us!!! before the deadline
The template of the report can be downloaded from www.manoonpong.com/ALCLecture/Report_Template.docx
www.manoonpong.com/ALCLecture/Report_Template.pdf
We will not evaluate the report that does not follow the given format !!
IF there are similar reports, both will fail!
About the course
Recommended books:
1) How the Body Shapes the Way We Think: A New View of Intelligence, by Rolf Pfeifer , Josh Bongard
2) Neural Preprocessing and Control of Reactive Walking Machines: Towards Versatile Artificial Perception-Action Systems (Cognitive Technologies), by Poramate Manoonpong
3) Adaptive Neural Control of Walking Robots by M.J. Randall
4) Principles of Insect Locomotion by H. Cruse, V. Dürr, M. Schilling, J. Schmitz
5) Neuronal Control of Locomotion: From Mollusc to Man (Oxford Neuroscience S) by G. N. Orlovsky, T. G. Deliagina and S. Grillner
About the course
4
Book:
Neural Computation in Embodied Closed-Loop Systems for the Generation of Complex Behavior: From Biology to Technology. Front. Neurorobot.
5
What have you learned?
What have you learned?
What have you learned?
9
Contents
1: Embodied AI: Design Principles of Intelligence (17.10.2020, 8:00-12:00)
2: Locomotion Principles in Animals (24.10.2020, 8:00-12:00)
3: Biomechanics and Neural Mechanisms for Adaptive Locomotion of Robots (31.10.2020, 8:00-12:00)
4: Neural Locomotion Control I (07.11.2020, 8:00-10:00) + CPG implementation on robot simulation for locomotion control (Practice, 10:00-12:00)
5: Neural Locomotion Control II (14.11.2020, 8:00-10:00) + Premotor neuron (VRN) implementation on robot simulation for directional control (Practice, 10:00-12:00)
6: Neural Sensory Preprocessing (21.11.2020, 8:00-10:00) + Neural preprocessing implementation on robot simulation for autonomous obstacle avoidance (Practice, 10:00-12:00)
7: Learning and Adaptation for Adaptive Locomotion I (28.11.2020, 8:00-10:00) + Robot learning implementation on simulation for adaptive behavior (Practice, 10:00-12:00)
8: Learning and Adaptation for Adaptive Locomotion II (05.12.2020, 8:00-10:00) + Robot learning for autonomous behavior generation (Practice, 10:00-12:00)
From Dec 6th onward, project and report
Week1: Embodied AI: Design Principles of Intelligence
→Thinking and body cannot be separated!
→ Complete Embodied Agent• Embodied→ Physical body • Autonomous→Without human control• Self-sufficient→ Sustain itself• Situated→ Sense & learn about the environment
→ Properties of Complete Embodied Agents
• Obey the laws of physics• Generate sensory stimulation• Effect the environment (interaction)• Perform morphological computation• Behave as dynamical systems (behaviors as
attractors!)
• Principle I: The Three-Constituents Principle(Structural design, Desired behaviors, Ecological niche)
• Principle II: The Complete-Agent Principle (Sensors & actuators, neural & morphological computations)
• Principle III: Cheap Design (Exploiting body to simplify control)
• Principle IV: Redundancy(Different sensor modalities)
• Principle V: Sensory-Motor Coordination (Correlation between sensory inputs & motor outputs)
• Principle VI: Ecological Balance (Balance between sensors, motors, materials, control)
Principle VII: Parallel, Loosely Coupled Processes(Distributed control with sensory feedback)
• Principle VIII: Value (Objective function for learning & adaptation)
Embodiment – Theoretical scheme (Pfeifer et al. 2007).
11
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials
Neural
mechanisms• Control
• Memory
• Learning
Dickinson et al., 2000
contact
musclelength, force
“Motor Intelligence”
• Joint coordination• Leg coordination
Week2: Locomotion Principles in Animals
12
Artificial systems (Robots)
Biomechanics &
Sensors (Body)• Structures
• Actuators / Muscles
• Materials
Neural
mechanisms• Control
• Memory
• Learning
“Motor Intelligence”
• Joint coordination• Leg coordination
CWRU
Ant-robot
HECTOR
RHEX
LAURON
HECTOR
AMOS
13
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials
How many motors do you need to build a hexapod robot?
14
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials 1-2 DOFs in total
The Dynamic Autonomous Sprawled Hexapod, DASH (Birkmeyer et al., 2009)
University of California
16
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials
Rhex, Saranli et al. 2001, 2004
1 DOF legs & non-segmented body(6 DOFs in total)
Additional DOF for adjustment
Galloway et al., 2011
17
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials 1 DOF legs & segmented body(7 DOFs in total)
http://biorobots.case.edu/projects/whegs/whegs-ii/
Schroer et al., 2004
18
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials
Robot I, CWRU
2 DOFs legs & non-segmented body(12 DOFs in total)
Fischer et al., 2005
http://www-cdr.stanford.edu/biomimetics/documents/sprawl/
Sprawlita hexapedal robot
19
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials 3 DOFs legs & non-segmented body(18 DOFs in total)
LAURON V, KITBILL-Ant-p robot, CWRU
Hector robot, Bielefeld (Schneider et al., 2014)
20
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials 3 DOFs legs & segmented body( >18 DOFs in total)
AMOSII robot, SDU&BCCN
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Billeschou, P., Bijma, N.N., Larsen, L.B., Gorb, S.N., Larsen, J.C., Manoonpong, P.
(2020) Framework for Developing Bio-inspired Morphologies for Walking Robots,Appl.
Sci. 2020, 10(19), 6986; https://doi.org/10.3390/app10196986
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Billeschou, P., Bijma, N.N., Larsen, L.B., Gorb, S.N., Larsen, J.C., Manoonpong, P.
(2020) Framework for Developing Bio-inspired Morphologies for Walking Robots,Appl.
Sci. 2020, 10(19), 6986; https://doi.org/10.3390/app10196986
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials investigate
Artificial systems
Biomechanics &
Sensors (Body)• Structures (summary!)
• Muscles
• Materials 1 DOF legs(6 DOFs in total)
1 DOF legs & segmented body(7 DOFs in total)
2 DOFs legs(12 DOFs in total)
3 DOFs legs (18 DOFs in total)
3 DOFs legs & segmented body( >18 DOFs in total)
Walking, Climbing
Walking, Climbing,Manipulating
(1-2 DOFs in total)
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
•Simple, cheap
•Revolve freely 360 degrees
•Easy to control
•H-bridge
•Using shaft encoders for feedback
•DC motor+gearing+feedbackcontrol loop circuit (position control)
•Expensive
•Position control (angle)
•Pulse Width Modulation (PWM)
•Mostly, 0-90, 0-180 degrees
•Position control
•No feedback used
•Control by impulses (switching
sequence)
•Worse weight/performance
ratio than DC, servo motors
DC motors Servo motors Stepper motors
Actuators (Rotary motion)
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Actuators (Linear motion)
•Linear movement•High forces without gears•Controlled by a hydraulic pump (heavy)•Noise & smoke of engine
•Linear movement•Fast response•Simple to control•Pull/push
•Linear movement•High forces•Power from compressed air (heavy)•Fast on/off
•Linear movement•Similar tostandard servo motor•PWM •Position control
Linear servo Pneumatic HydraulicSolenoids
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
Ham et al., 2009
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
”But Muscles are not just stiff but also compliant!”
ETH, Zurich
Ham et al., 2009
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
MACCEPA (Mechanically Adjustable Compliance and Controllable Equilibrium Position Actuator)
Vanderborght et al., 2009
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
Antagonistic setup of two SEA
AMASC
Hurst et al., 2004
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
Antagonistic setup of two SEA
AMASC
Hurst et al., 2004
Migliore et al., 2005
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant Actuators
Compact series elastic actuator
Tsagarakis et al., 2009
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Pneumatic artificial muscles
•Flexible bladder surrounded by nylon cord•Power from compressed air (heavy)•Fast response•Nonlinearity
Shape memory alloy SMA
•Copper-zinc-alu-nikel, Copper-alu-nickel, Nickel-titanium alloys•Changing shape when energy is applied toOr removed from them•Light weight•Low force output, poor fatigue properties
1968 McKibben
Compliant Actuators
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant ActuatorsActuator
Spring
Load
”Hardware”
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Compliant ActuatorsActuator
Spring
Load
Actuator
Virtual Spring
LoadStiff Actuators with Software Control
”Software”
”Hardware”
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Active compliance using force/torque sensing at each joint (force control)
https://www.youtube.com/watch?v=r6mrNnIamKw
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Standard servomotor
(position control)
without angle feedback
+ = Variable
compliancesMusclemodel
M1M2
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Musclemodel
Virtual Agonist-Antagonist Muscle (i.e., M1 and M2).
It consists of parallel and contractile elements (i.e., PE(1,2) and CE(1,2)).
PE(1,2) for compliance (passive force)
CE(1,2) for actuation (active force)
M1M2
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Musclemodel
Virtual Agonist-Antagonist Muscle (i.e., M1 and M2).
It consists of parallel and contractile elements (i.e., PE(1,2) and CE(1,2)).
PE(1,2) for compliance (passive force)
CE(1,2) for actuation (active force)
Passive forceActive forceExternal force
M1M2
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Musclemodel
Virtual Agonist-Antagonist Muscle (i.e., M1 and M2).
It consists of parallel and contractile elements (i.e., PE(1,2) and CE(1,2)).
PE(1,2) for compliance (passive force)
CE(1,2) for actuation (active force)
Passive forceActive forceExternal force
M1M2
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Musclemodel
Virtual Agonist-Antagonist Muscle (i.e., M1 and M2).
It consists of parallel and contractile elements (i.e., PE(1,2) and CE(1,2)).
PE(1,2) for compliance (passive force)
CE(1,2) for actuation (active force)
Passive forceActive forceExternal force
M1M2
“Runge–Kutta
method”
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Musclemodel
Virtual Agonist-Antagonist Muscle (i.e., M1 and M2).
It consists of parallel and contractile elements (i.e., PE(1,2) and CE(1,2)).
PE(1,2) for compliance (passive force)
CE(1,2) for actuation (active force)
Passive forceActive forceExternal force
M1M2
“Runge–Kutta
method”
Tuning K or D for
variable compliance!
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
0
Force sensor
Two-jointed legs
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Softer joints (i.e., D = 0.001, K = 0.8)→ bouncing
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Stiffer joints (i.e., D = 0.1, K = 0.8)→difficult to push
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Intermediately soft joints (i.e., D = 0.01, K = 0.8)
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Intermediately soft joints (i.e., D = 0.01, K = 0.8)
Artificial systems
Biomechanics &
Sensors (Body)• Muscles
Stiff Actuators with Software Control
Virtual muscle model using force at the foot of a leg (position control)
Intermediately soft joints (i.e., D = 0.01, K = 0.8)
5.72kg Load
Xiong et al., Industrial Robot, 2014
Artificial systems
Biomechanics &
Sensors (Body)• Muscles (summary!)
Spring or damper
Spring or damper
Artificial systems
Biomechanics &
Sensors (Body)• Materials
Steel, Aluminium
Carbon-fiber-reinforced polymer
Rigid
Plastic
57
Artificial systems
Biomechanics &
Sensors (Body)• Materials
Steel, Aluminium
Carbon-fiber-reinforced polymer
Rigid
Plastic
Spring
Tendon
58
Artificial systems
Biomechanics &
Sensors (Body)• Materials
Steel, Aluminium
Carbon-fiber-reinforced polymer
Rigid Soft
Plastic
SpringRubber
SiliconeECOFLEX
Tendon
Pourable Polyurethane
59
Artificial systems
Biomechanics &
Sensors (Body)• Materials
Shepherd et al., 2011
Harvard University
60
Artificial systems
Biomechanics &
Sensors (Body)• Materials
Shape Deposition Manufacturing (SDM)
Sprawlita hexapedal robot (Cham et al., 2002)
October 2013 61
30o
4X
Incline walking (rough surface)
Reactive posture control (Foot & IMU sensors)
25°. Roennau et al., IEEE IROS, 2014
4X
Control:New walking behavior
Incline walking (rough surface)
30o
Morphological
computation:Structure/material
30o
4X
Control:New walking behavior
Morphological
computation:Structure/material
Incline walking (rough surface)
Adhesion
Mechanical
interlocking
Smooth surface
Rough surface
Spenko et al., J. Field Robot., 2008
Kimet al., IEEE Trans. Robot., 2008
Stanford University
30o
4X
Control:New walking behavior
Morphological
computation:Structure/material
Incline walking (rough surface)
Mechanical
interlockingRough surface
Spenko et al., J. Field Robot., 2008
FOOT RASP CALLUS dead
skin remover
Manoonpong et al., Scientific Reports-Nature Journal, 2016, In collaboration with Kiel University (Prof. Gorb)
4X
Manoonpong et al., Scientific Reports-Nature Journal, 2016, In collaboration with Kiel University (Prof. Gorb)
Incline walking (rough surface)
Frictional isotropy
(Uniform spikes)
30o
4X
Frictional anisotropy
(Shark skin, Sloped spikes)
Manoonpong et al., Scientific Reports-Nature Journal, 2016, In collaboration with Kiel University (Prof. Gorb)
Incline walking (rough surface)
Rough surface
30o
4XManoonpong et al., Scientific Reports-Nature Journal, 2016, In collaboration with Kiel University (Prof. Gorb)
Incline walking (rough surface)
SR = P/mgv
Locomotion efficiency of AMOS
with and without shark skin
Manoonpong et al., Scientific Reports-Nature Journal, 2016, In collaboration with Kiel University (Prof. Gorb)
Sharkskin shoes
Manoonpong et al., Scientific Reports-Nature Journal, 2016, In collaboration with Kiel University (Prof. Gorb)
Sharkskin shoesMicro scale
4X
4X17 deg
17 deg
78
Artificial systems
Biomechanics &
Sensors (Body)• Materials
Not only soft materials are good but also different surfaces for variousfrictions are required!
79
Artificial systems
Biomechanics &
Sensors (Body)• Materials (summary!)
• Rigid• Soft• Anisotropic friction
80
Artificial systems
Biomechanics &
Sensors (Body)
• Structures →
• Muscles →
• Materials →
Segmented legs & body
Actuation & compliance
Rigid, soft, anisotropic friction
“Virtual muscles”
“Shark skin”
“Rubber&
spring”
“Backbone” “3 Jointed legs ”
AMOSII (19 motors)
81
Artificial systems
Biomechanics &
Sensors (Body)
• Structures →
• Muscles →
• Materials →
Segmented legs & body
Actuation & compliance
Rigid, soft, anisotropic friction
“Virtual muscles”
“Shark skin”
“Rubber&
spring”
“Backbone” “3 Jointed legs ”
AMOSII (19 motors)
October 2013 82
Biomechanics &
Sensors (Body)
Neural
mechanisms
MemoryLearningControl
StructureMusclesMaterials
Biomechanical
development
Bio-inspired
walking robotsNeural mechanisms
(modular structure)
Structure: ?Muscles: ?Material: ?
Artificial systems
Short Summary
Walking animals
(Insects)
Biomechanics &
Sensors (Body)
Neural
mechanisms
MemoryLearningControl
StructureMusclesMaterials
Biomechanical
development
Bio-inspired
walking robotsNeural mechanisms
(modular structure)
Structure: ?Muscles: ?Material: ?
Short Summary
Walking animals
(Insects)
Artificial systems
85
Biological systems
Biomechanics &
Sensors (Body)• Sensors
Proprioceptive sensing (sensing internal environment)
Exteroceptive sensing (sening the external environment)
86
Artificial systems
Biomechanics &
Sensors (Body)• Sensors
•Poteniometers (Pot), shaft encoder (En)
•Determining rotational motion (joint angle)
•Variable resistor (Pot)
•Built in a servo motor (Pot)
•Analog (Pot, En) / Digital (En) output
•Passive sensors
•Accelerometer: measuring the acceleration
•Gyroscope: measuring the rotational change
•Inclenometer: measuring the absolute angle
•Ball switch: measuring the rotational change
•Balancing
•Analog, PWM, digital, output
•Passive sensors
Position sensing (leg joints) Orientation sensing (body)
Proprioceptive sensing (sensing internal environment)
Exteroceptive sensing (sening the external environment)
87
Artificial systems
Biomechanics &
Sensors (Body)• Sensors
Proprioceptive sensing (sensing internal environment)
Exteroceptive sensing (sening the external environment)
•Stain gauge, force sensing resistors (FSR)
•Determining loading being carried by the
machines, foot contact event
•Variable resistance
•Analog output
•Passive sensors
•Current sensor, voltage sensor
•Determining current, voltage
•Analog output
•Passive sensors
Load sensing (feet) Power management
88
Artificial systems
Biomechanics &
Sensors (Body)• Sensors
Proprioceptive sensing (sensing internal environment)
Exteroceptive sensing (sening the external environment)
•IR sensor, sonar sensor,
laser scanner
•Measuring distance
•Analog/ digital output
•Active sensors
•Camera, most complex sensors
used
•Image processing, e.g., object
recognition
•Analog/digital output (USB)
•Passive sensor
•Photoresistor or light
dependent resistor
•Detecting light source
•Variable resistance
•Analog output
•Passive sensor
Proximity sensing Vision Light sensing
89
Artificial systems
Biomechanics &
Sensors (Body)• Sensors
Proprioceptive sensing (sensing internal environment)
Exteroceptive sensing (sening the external environment)
•Microphone
•Sound localization
•Analog output
•Passive sensor
•Direct contact, bumpers,
Artificial skin, antenna
•Measuring distance, tactile
sensing
•Analog/digital output
•Passive sensor
•Microphone + whisker
•Detecting sound, wind
tactile sensing
•Analog output
•Passive sensor
Audition Touch Auditory-tactile-wind sensing
90
Artificial systems
Biomechanics &
Sensors (Body)• Sensors
Proprioceptive sensing (sensing internal environment)
Exteroceptive sensing (sening the external environment)
Compass (internal, passive)Temperature sensor (internal/external, passive)
Olfaction (external, passive)
GPS (external, active)
Vibration (internal/external, passive)
Biomechanics &
Sensors (Body)
Neural
mechanisms
MemoryLearningControl
StructureMusclesMaterials
Biomechanical
development & sensorsBio-inspired
walking robotsNeural mechanisms
(modular structure)
Structure: ?Muscles: ?Material: ?
Short Summary
Walking animals
(Insects)
Artificial systems
Biomechanics &
Sensors (Body)• Structures
• Muscles
• Materials
Neural
mechanisms• Control
• Memory
• Learning
“Motor Intelligence”
• Joint coordination• Leg coordination
Locomotion in Robots
Locomotion control of walking robots
92
The problems of legged locomotion control
Coordinating all the degrees-of-freedom of the robot: finding the right
• Frequency f=1/T• Phase ϕ,
• Amplitude A
• Signal shape
Ijspeert, EPFL Locomotion control
93
The problems of legged locomotion control
•Need to modify the gait for different speeds and directions
•Need to keep balance
•Need to take advantage of the robot’s dynamics
•Obstacle avoidance
• Adapting to perturbations or different terrains
•Optimizing the gaits (finding the fastest, most efficient,...)
• Dealing with lesions, and changes in body properties
Ijspeert, EPFL Locomotion control
94
Minimal ingredients for locomotion control
1. A trajectory planner:
• For producing the different trajectories that each degree of freedom (DOF) has to follow.
• Trajectory = set of joint angles over time
2. A feedback PID-controller:
• For producing the torques (in the motors) necessary to follow a specific trajectory.
Ijspeert, EPFL Locomotion control95
Minimalistic control diagram (block diagram)
(Desired joint angle)
(Actual joint angle)
(Joint angle command)
Ijspeert, EPFL Locomotion control
96
Locomotion control
Forward/inverse kinematics, balance
control, impedance/stiffness control
Semini et al.,
Int J Rob Res., 2015
Boston Dynamics, USA, 2018
IEEE SPECTRUM 2019,
Zhejiang University
Different approaches to legged locomotion control in current robots
I. NEED kinematic
models
Engineering control
Forward/inverse kinematics, balance
control, impedance/stiffness control
Semini et al.,
Int J Rob Res., 2015
Boston Dynamics, USA, 2018
IEEE SPECTRUM 2019,
Zhejiang University
Different approaches to legged locomotion control in current robots
I. NEED kinematic
models
Engineering control
(Foot tip)
Model-based control (traditional control engineering): Forward/Inverse kinematics
(Foot tip)
Model-based control (traditional control engineering): Forward/Inverse kinematics
d1
a2 a3
Model-based control (traditional control engineering): Forward/Inverse kinematics
BaseEnd effector
End effector
2
1
3
2
1
3
d1
a2a3
Model-based control (traditional control engineering): Forward/Inverse kinematics
C1,2,3 = cos 1,2,3; C23 = cos(2+3); S1,2,3 = sin 1,2,3; S23 = sin(2+3)
Px
Py
Pz
d1
a2 a3
End effector
2
1
3
Model-based control (traditional control engineering): Forward/Inverse kinematics
C1,2,3 = cos 1,2,3; C23 = cos(2+3); S1,2,3 = sin 1,2,3; S23 = sin(2+3)
Px
Py
Pz
q1= 1= atan2(Py, Px);
COSZETA3=((Px*Px)+(Py*Py)+((Pz-d1)*(Pz-d1))-(a3*a3)-(a2*a2))/(2*a2*a3);
q3= 3 = atan2(sqrt(1-COSZETA3*COSZETA3), COSZETA3);
q2= 2 =-atan2((Pz-m_d1), sqrt(Px*Px+Py*Py))-atan2(m_a3*sin(q3),m_a2+m_a3*cos(q3));
d1
a2 a3
End effector
2
1
3
Model-based control (traditional control engineering): Forward/Inverse kinematics
C1,2,3 = cos 1,2,3; C23 = cos(2+3); S1,2,3 = sin 1,2,3; S23 = sin(2+3)
Px
Py
Pz
q1= 1= atan2(Py, Px);
COSZETA3=((Px*Px)+(Py*Py)+((Pz-d1)*(Pz-d1))-(a3*a3)-(a2*a2))/(2*a2*a3);
q3= 3 = atan2(sqrt(1-COSZETA3*COSZETA3), COSZETA3);
q2= 2 =-atan2((Pz-m_d1), sqrt(Px*Px+Py*Py))-atan2(m_a3*sin(q3),m_a2+m_a3*cos(q3));
Inputs (Px, Py, Pz)
d1
a2 a3
End effector
2
1
3
Trajectory generation (Cartesian space)
R3
R2
R1
L3
L2
L1
”Tripod gait”
Px = 0.0
x
Y
Z
105
Trajectory generation (Cartesian space)
R3
R2
R1
L3
L2
L1
”Tripod gait”
Px = 0.0
x
Y
Z
Manual design!!!106
Trajectory generation (Cartesian space)
R3
R2
R1
L3
L2
L1
”Tripod gait”
Px = 0.0
x
Y
Z
Recoding animal walking!!!107
Stride length -> 10 cm.
Pz
Py
Px = 0.0
Trajectory generation (Cartesian space)
Inverse kinematics
ST SW ST
2
1
3
1
2
3
108
109
Forward/inverse kinematics, balance
control, impedance/stiffness control
Prior knowledge map & offline trial-
and-error learning with approx.
40 million iterations
Offline reinforcement learning
with approx. 2600 iterations
Semini et al.,
Int J Rob Res., 2015
Hwangbo et al., Science Robotics, 2019
ETH, Switzerland
Boston Dynamics, USA, 2018
IEEE SPECTRUM 2019,
Zhejiang University
Different approaches to legged locomotion control in current robots
Machine learning
Cully et al., Nature, 2015
I. NEED kinematic
modelsII. NEED long learning time
(mins to hours/days in Sim)
Engineering control
Evolutionary algorithm
Approx. 250,000 internal modelsimulations
Evolving robot control for locomotion
Bongard, J., Zykov, V., Lipson, H. (2006). Resilient machines through continuous self-modeling. Science, 314: 1118-1121.
Approx. 250,000 internal modelsimulations
Forward/inverse kinematics, balance
control, impedance/stiffness control
Prior knowledge map & offline trial-
and-error learning with approx.
40 million iterations
Offline reinforcement learning
with approx. 2600 iterations
Semini et al.,
Int J Rob Res., 2015
Hwangbo et al., Science Robotics, 2019
ETH, Switzerland
Boston Dynamics, USA, 2018
IEEE SPECTRUM 2019,
Zhejiang University
Different approaches to legged locomotion control in current robots
Machine learning
Cully et al., Nature, 2015
I. NEED kinematic
modelsII. NEED long learning time
(mins to hours/days in Sim)
Engineering control
Evolutionary algorithm
Approx. 250,000 internal modelsimulations
Forward/inverse kinematics, balance
control, impedance/stiffness control
Prior knowledge map & offline trial-
and-error learning with approx.
40 million iterations
Offline reinforcement learning
with approx. 2600 iterations
Semini et al.,
Int J Rob Res., 2015
Hwangbo et al., Science Robotics, 2019
ETH, Switzerland
Boston Dynamics, USA, 2018
IEEE SPECTRUM 2019,
Zhejiang University
Different approaches to legged locomotion control in current robots
Machine learning
Cully et al., Nature, 2015
I. NEED kinematic
modelsII. NEED long learning time
(mins to hours/days in Sim)
Engineering control
Forward/inverse kinematics, balance
control, impedance/stiffness control
Prior knowledge map & offline trial-
and-error learning with approx.
40 million iterations
Offline reinforcement learning
with approx. 2600 iterations
Semini et al.,
Int J Rob Res., 2015
Hwangbo et al., Science Robotics, 2019
ETH, Switzerland
Boston Dynamics, USA, 2018
IEEE SPECTRUM 2019,
Zhejiang University
Different approaches to legged locomotion control in current robots
Machine learning
Cully et al., Nature, 2015
I. NEED kinematic
modelsII. NEED long learning time
(mins to hours/days in Sim)
Engineering control
CPGs-based control with
predefined coordination
Bio-inspired control
Karakasiliotis et al., J. R. Soc. Interface, 2016
EPFL, Switzerland
Schilling et al., Biol Cybern., 2013 (EU project)
Reflex-based control (Purely
sensor driven with predefined
coordination
III. CPG, Reflex,
Neural Control …
Different approaches to legged locomotion control in current robots
CPGs-based control with
predefined coordination
Bio-inspired control
Karakasiliotis et al., J. R. Soc. Interface, 2016
EPFL, Switzerland
Schilling et al., Biol Cybern., 2013 (EU project)
Reflex-based control (Purely
sensor driven with predefined
coordination
III. CPG, Reflex,
Neural Control …
Central Pattern Generators (CPGs)
Ryczko D., Simon A., Ijspeert AJ. (2020) Walking with Salamanders: From Molecules to Biorobotics, Trends in Neurosciences
Different approaches to legged locomotion control in current robots
CPGs-based control with
predefined coordination
Bio-inspired control
Karakasiliotis et al., J. R. Soc. Interface, 2016
EPFL, Switzerland
Schilling et al., Biol Cybern., 2013 (EU project)
Reflex-based control (Purely
sensor driven with predefined
coordination
III. CPG, Reflex,
Neural Control …
Neural mechanisms underlying locomotion
Neural mechanisms underlying locomotion
Neural mechanisms underlying locomotion
Neural mechanisms underlying locomotion
Biological neural networksArtificial neural networks
Neural mechanisms underlying locomotion
Biological neural networksArtificial neural networks
“From Number
To Behavior !”1110001010101
00100010011
110001111
0110010
11100
00010
1101
11
Biomechanical
development & sensorsBio-inspired
walking robots
Structure: ?Muscles: ?Material: ?
Summary
Biomechanical
development & sensorsBio-inspired
walking robotsLocomotion control
Structure: ?Muscles: ?Material: ?
Summary
Machine learning
Engineering control
Bio-inspired control
Biomechanical
development & sensorsBio-inspired
walking robotsLocomotion control
Structure: ?Muscles: ?Material: ?
Summary
Machine learning
Engineering control
Bio-inspired control
Next week
NEXT WEEK: Neural Locomotion Control ISe
nso
rs
Mo
tors
Memory guidance
Goal-directed navigation
Neural learning (Online)
Neural locomotion control
Learning & Adaptation
Control
• Central Pattern Generator (CPG)
• Frequency adaptation
• Robot simulation !!! Please prepare your notebook and install the robot simulation software before next time!!, classroom at NUAA Ming Palace Campus
Reading Materials of Today!
1) Rolf Pfeifer, Max Lungarella, Fumiya Iida (2012) The challenges ahead for bio-inspired 'soft' robotics, Communications of the ACM, Volume 55 Issue 11, November 2012 Pages 76-87
2) Ham, R.; Sugar, T.G. ; Vanderborght, B. ; Hollander, K.W. ; Lefeber, D. (2009) Compliant actuator designs, IEEE Robotics & Automation Magazine, pp. 81-94
3) Sangbae Kim, Cecilia Laschi, Barry Trimmer(2013) Soft robotics: a bioinspired evolution in robotics, Trends in Biotechnology, Volume 31, Issue 5, May 2013, Pages 287–294
4) Xiaodong Zhou and Shusheng Bi (2012) A survey of bio-inspired compliant legged robot designs, Bioinspir. Biomim. 7, 041001 (20pp)
5) Xiong, X.; Wörgötter, F.; Manoonpong, P.(2014) Virtual Agonist-antagonist Mechanisms Produce Biological Muscle-like Functions: An Application for Robot Joint Control. Industrial Robot: An International Journal, Vol. 41 Iss: 4
147
Exercises & project work
1) Robot simulation (C++, gorobots_edu)
1. Assuming you have a clean install of Ubuntu 18!!https://releases.ubuntu.com/bionic/
2. Please follow this link for installing the simulationhttp://manoonpong.com/MOROCO/lpz_guide.txt
Appendix: Muscle model
Passive force
Passive force
Active force
Active force
External forcespring damper
spring damper
Passive forceActive forceExternal force
spring damperNeural activation
Neural activation
M1M2