Upload
vodung
View
223
Download
1
Embed Size (px)
Citation preview
Improvement of the Cheetah Locomotion ControlMaster Project - Final Presentation
15th February 2010
Student : Alexandre TuleuSupervisor : Alexander SpröwitzProfessor : Auke Jan Ijspeert
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Presentation of the Cheetah
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Presentation of the Cheetah
• Light weighted, biollogically inspired small quadruped robot
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Presentation of the Cheetah
• Light weighted, biollogically inspired small quadruped robot
• Pantograph, compliant legs
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Presentation of the Cheetah
• Light weighted, biollogically inspired small quadruped robot
• Pantograph, compliant legs
• Use CPG and Optimization Process for offline learning
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Goals of the project
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Goals of the project
• Improve the locomotion control in term of efficiency, Robustness and controllability
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Goals of the project
• Improve the locomotion control in term of efficiency, Robustness and controllability
• Model mechanic improvement, and measure their effect on the locomotion behavior.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Table of contents
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Table of contents
• Self-stabilization strategy in Nature
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Table of contents
• Self-stabilization strategy in Nature
• Timeline of the project
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Table of contents
• Self-stabilization strategy in Nature
• Timeline of the project
• Update of the webots model
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Table of contents
• Self-stabilization strategy in Nature
• Timeline of the project
• Update of the webots model
• CPG Design
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Table of contents
• Self-stabilization strategy in Nature
• Timeline of the project
• Update of the webots model
• CPG Design
• Optimization Design & result
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Three Self-Stabilization Principles
Introduction Self-Stabilization Strategy Model Update Control Design Optimization
[Seyfarth03] Swing Leg Retraction: a simple control model for stable running, André Seyfarth and Hartmut Geyer, Journal of experimental Zoology 206:2547-2555 (2003)[Daley09] The Role of Intrinsic muscle mechanics in the neuromuscular control of stable running in the guinea fowl, Journal of Physiology 587.11 (2009) pp 2693-2707Video Extract : Youtube
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Three Self-Stabilization Principles
Passive capacity to stabilize in open loop process (no sensory feed back as a response to internal or external disturbance)
Self-Stabilization
Introduction Self-Stabilization Strategy Model Update Control Design Optimization
[Seyfarth03] Swing Leg Retraction: a simple control model for stable running, André Seyfarth and Hartmut Geyer, Journal of experimental Zoology 206:2547-2555 (2003)[Daley09] The Role of Intrinsic muscle mechanics in the neuromuscular control of stable running in the guinea fowl, Journal of Physiology 587.11 (2009) pp 2693-2707Video Extract : Youtube
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Three Self-Stabilization Principles
• Leg Stiffening
Passive capacity to stabilize in open loop process (no sensory feed back as a response to internal or external disturbance)
Self-Stabilization
Introduction Self-Stabilization Strategy Model Update Control Design Optimization
[Seyfarth03] Swing Leg Retraction: a simple control model for stable running, André Seyfarth and Hartmut Geyer, Journal of experimental Zoology 206:2547-2555 (2003)[Daley09] The Role of Intrinsic muscle mechanics in the neuromuscular control of stable running in the guinea fowl, Journal of Physiology 587.11 (2009) pp 2693-2707Video Extract : Youtube
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Three Self-Stabilization Principles
• Leg Stiffening
• Leg Retraction [Seyfarth03]
Passive capacity to stabilize in open loop process (no sensory feed back as a response to internal or external disturbance)
Self-Stabilization
Introduction Self-Stabilization Strategy Model Update Control Design Optimization
[Seyfarth03] Swing Leg Retraction: a simple control model for stable running, André Seyfarth and Hartmut Geyer, Journal of experimental Zoology 206:2547-2555 (2003)[Daley09] The Role of Intrinsic muscle mechanics in the neuromuscular control of stable running in the guinea fowl, Journal of Physiology 587.11 (2009) pp 2693-2707Video Extract : Youtube
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Three Self-Stabilization Principles
• Leg Stiffening
• Leg Retraction [Seyfarth03]
• Leg Extension[Daley09]
Passive capacity to stabilize in open loop process (no sensory feed back as a response to internal or external disturbance)
Self-Stabilization
Introduction Self-Stabilization Strategy Model Update Control Design Optimization
[Seyfarth03] Swing Leg Retraction: a simple control model for stable running, André Seyfarth and Hartmut Geyer, Journal of experimental Zoology 206:2547-2555 (2003)[Daley09] The Role of Intrinsic muscle mechanics in the neuromuscular control of stable running in the guinea fowl, Journal of Physiology 587.11 (2009) pp 2693-2707Video Extract : Youtube
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Timeline
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Model Update Design of the control Testing Mechanical Improvement
Fix Webots Model
Foot trajectory definition
Prepare optimization framework
Gait Optimization
New Foot
New Scapula Joint
Sipnal Cord
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Timeline
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Model Update Design of the control Testing Mechanical Improvement
Fix Webots Model
Foot trajectory definition
Prepare optimization framework
Gait Optimization
New Foot
New Scapula Joint
Sipnal Cord
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Timeline
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Model Update Design of the control Testing Mechanical Improvement
Fix Webots Model
Foot trajectory definition
Prepare optimization framework
Gait Optimization
New Foot
New Scapula Joint
Sipnal Cord
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Modelisation of the Knee
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Modelisation of the Knee
• Enhanced Model of the Knee Servo motor implemented
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Modelisation of the Knee
• Enhanced Model of the Knee Servo motor implemented
• Bowden cable modeled as an assymetrical spring
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Modelisation of the Knee
• Enhanced Model of the Knee Servo motor implemented
• Bowden cable modeled as an assymetrical spring
• Further Improvement : decrease timestep integration with a loop
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
ODE unrealistic behavior
Previous solution : tuning the ODE CFM parameter.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
ODE unrealistic behavior
Previous solution : tuning the ODE CFM parameter.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Releasing the overconstraint
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Releasing the overconstraint
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Releasing the overconstraint
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Ludovic Righetti CPG
Part 3
Implementation of the locomotioncontrol
3.1 Central Pattern Generator defintion and extensionFor the control of the Cheetah a Central Pattern Generator, previously designed by Ludovic
Righetti[RI08] has been used. This CPG is able to generate up to four gait found in quadrupedal
(walk, trot, pace and bound), and features sensory feedback, that synchronizes the command with
the state of the foot (stance and swing phases). After a presentation of this CPG, a modification is
proposed to introduce the leg retraction strategy.
3.1.1 A Central Pattern Generator for quadrupedal walking3.1.1.a General presentation
The CPG proposed by [RI08], is based on a modified Hopf oscillator. There are four neurons
which have a state (xi,yi)
xi = α(µ− r2)xi−ωiyi (3.1)
yi = β (µ− r2)yi +ωixi + ∑j �=i
ki jy j (3.2)
ωi =ωst
1+ eby +ωsw
1+ e−by (3.3)
According to [RI08], this oscillator exhibit a limit cycle (see figure 3.1) , which is the circle of
radiusõ in the phase space (x,y). [RI08] shows that the choose of the appropriate coupling matrix
(ki, j) for four oscillators, leads them to phase lock. The resulting difference of phase between the
oscillator depends on symmetries between block of the chosen matrix (see figure 3.2 as an example).
[RI08] also gives 4 matrices for each of the following gait : walk, trot pace and bound.
Figure 3.1 – Representation of the attraction field of one oscillator. from [RI08]
26
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Ludovic Righetti CPG
• Can generate several gait : walk trot pace bound.
Part 3
Implementation of the locomotioncontrol
3.1 Central Pattern Generator defintion and extensionFor the control of the Cheetah a Central Pattern Generator, previously designed by Ludovic
Righetti[RI08] has been used. This CPG is able to generate up to four gait found in quadrupedal
(walk, trot, pace and bound), and features sensory feedback, that synchronizes the command with
the state of the foot (stance and swing phases). After a presentation of this CPG, a modification is
proposed to introduce the leg retraction strategy.
3.1.1 A Central Pattern Generator for quadrupedal walking3.1.1.a General presentation
The CPG proposed by [RI08], is based on a modified Hopf oscillator. There are four neurons
which have a state (xi,yi)
xi = α(µ− r2)xi−ωiyi (3.1)
yi = β (µ− r2)yi +ωixi + ∑j �=i
ki jy j (3.2)
ωi =ωst
1+ eby +ωsw
1+ e−by (3.3)
According to [RI08], this oscillator exhibit a limit cycle (see figure 3.1) , which is the circle of
radiusõ in the phase space (x,y). [RI08] shows that the choose of the appropriate coupling matrix
(ki, j) for four oscillators, leads them to phase lock. The resulting difference of phase between the
oscillator depends on symmetries between block of the chosen matrix (see figure 3.2 as an example).
[RI08] also gives 4 matrices for each of the following gait : walk, trot pace and bound.
Figure 3.1 – Representation of the attraction field of one oscillator. from [RI08]
26
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Ludovic Righetti CPG
• Can generate several gait : walk trot pace bound.
• Can modulate the duty ratio
Part 3
Implementation of the locomotioncontrol
3.1 Central Pattern Generator defintion and extensionFor the control of the Cheetah a Central Pattern Generator, previously designed by Ludovic
Righetti[RI08] has been used. This CPG is able to generate up to four gait found in quadrupedal
(walk, trot, pace and bound), and features sensory feedback, that synchronizes the command with
the state of the foot (stance and swing phases). After a presentation of this CPG, a modification is
proposed to introduce the leg retraction strategy.
3.1.1 A Central Pattern Generator for quadrupedal walking3.1.1.a General presentation
The CPG proposed by [RI08], is based on a modified Hopf oscillator. There are four neurons
which have a state (xi,yi)
xi = α(µ− r2)xi−ωiyi (3.1)
yi = β (µ− r2)yi +ωixi + ∑j �=i
ki jy j (3.2)
ωi =ωst
1+ eby +ωsw
1+ e−by (3.3)
According to [RI08], this oscillator exhibit a limit cycle (see figure 3.1) , which is the circle of
radiusõ in the phase space (x,y). [RI08] shows that the choose of the appropriate coupling matrix
(ki, j) for four oscillators, leads them to phase lock. The resulting difference of phase between the
oscillator depends on symmetries between block of the chosen matrix (see figure 3.2 as an example).
[RI08] also gives 4 matrices for each of the following gait : walk, trot pace and bound.
Figure 3.1 – Representation of the attraction field of one oscillator. from [RI08]
26
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Ludovic Righetti CPG
• Can generate several gait : walk trot pace bound.
• Can modulate the duty ratio
• Uses sensory feedback information
Part 3
Implementation of the locomotioncontrol
3.1 Central Pattern Generator defintion and extensionFor the control of the Cheetah a Central Pattern Generator, previously designed by Ludovic
Righetti[RI08] has been used. This CPG is able to generate up to four gait found in quadrupedal
(walk, trot, pace and bound), and features sensory feedback, that synchronizes the command with
the state of the foot (stance and swing phases). After a presentation of this CPG, a modification is
proposed to introduce the leg retraction strategy.
3.1.1 A Central Pattern Generator for quadrupedal walking3.1.1.a General presentation
The CPG proposed by [RI08], is based on a modified Hopf oscillator. There are four neurons
which have a state (xi,yi)
xi = α(µ− r2)xi−ωiyi (3.1)
yi = β (µ− r2)yi +ωixi + ∑j �=i
ki jy j (3.2)
ωi =ωst
1+ eby +ωsw
1+ e−by (3.3)
According to [RI08], this oscillator exhibit a limit cycle (see figure 3.1) , which is the circle of
radiusõ in the phase space (x,y). [RI08] shows that the choose of the appropriate coupling matrix
(ki, j) for four oscillators, leads them to phase lock. The resulting difference of phase between the
oscillator depends on symmetries between block of the chosen matrix (see figure 3.2 as an example).
[RI08] also gives 4 matrices for each of the following gait : walk, trot pace and bound.
Figure 3.1 – Representation of the attraction field of one oscillator. from [RI08]
26
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Adding leg retraction strategy to CPG
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Adding leg retraction strategy to CPG
• Need to have a non-zero horizontal velocity at touchdown
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Adding leg retraction strategy to CPG
• Need to have a non-zero horizontal velocity at touchdown
• Moves the swing to stance transition point.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Adding leg retraction strategy to CPG
• Need to have a non-zero horizontal velocity at touchdown
• Moves the swing to stance transition point.
• Just change the phase output.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Adding leg retraction strategy to CPG
• Need to have a non-zero horizontal velocity at touchdown
• Moves the swing to stance transition point.
• Just change the phase output.
• Transformation of the phase in polar coordinate.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
unchanged. The following properties for fρ are also required :
fρ(0) = 0 (3.20)fρ(−π) = ρ −π (3.21)
fρ(π) = ρ +π (3.22)f �ρ(0) = 1 (3.23)
f �ρ(−π) = f �ρ(π) (3.24)
• Equation (3.20) avoids any displacement of the stance to swing transition point.
• Equations (3.21) and (3.22) move the stance to swing transition point and are a necessarycondition, fρ to be a bijection.
• Equations (3.23) and (3.24) insure a smoothness of the solution
One can found a unique 4th order polynom that satisfies the conditions (3.20-3.24) :
fρ(θ) =−ρ.
�θπ
�4+2.ρ.
�θπ
�2+θ (3.25)
The result of one of these transformations is seen in figure 3.6.
(a) Original trajectories (b) Images of the trajectories (c) Transform of a uniform mesh
Figure 3.6 – Transformation of a few trajectories with the function fρ , for ρ = π4 .
One should also notice that all the conditions are not met, fρ to be a bijection. However ifρ ∈
�0, π
2�, this is always the case.
31
Here the choice of extending at maximum the leg at each gait cycle has been made. The height ofthe CoM can easily be controlled, by adding a new parameter νi, and so not extend the leg as itmaximum (ϕknee
i = 0. Once again, the following equations are introduce, to provide only smoothchanges of the two level parameters :
λi = kλ
�λi−λ des
i
�(3.17)
νi = kν�
νi−νdesi
�(3.18)
3.1.2 Addition of the leg retraction principle
The next goal is to implement the leg retraction principle into the CPG. More formally, wewant that at touchdown, the gesture of retraction has already begun, i.e. : that both x < 1 andx > 0. Ideally we also want to parameterize this amount of retraction, by for example specifyingx(ttouchdown). In [SGH03], this parameter is the constant rotational speed of the leg.
Figure 3.5 – Implementation of the leg retraction principle. The swing to stance transition point should bemoved of an angle ρ ≥ 0.
However, implementing the leg retraction principle this way is rather difficult. In [SGH03]the angle (to the vertical) of the leg at APEX is precisely controlled, and the rotational speed isconstant. The first predicate is difficult to implement in a robot, and the second one isn’t reachablewithin the servo limits. Cheetah is intended to be really fast, at the limit of the servo specification.So it couldn’t be expected, the servo not to have a delay to reach some specific speed.
Further more, the convergence properties of this CPG are rather complicated, and defining newlimit cycle would be difficult. Then it may be simpler, in a first test, just to transform the output ofthe oscillators. Then the new behavior won’t interfere with the properties of the dynamical system.The trajectories generated by the new output should also be changed, as little as possible, from theoriginal one. Therefore the limit cycle (unit circle) must not be changed.
In order to fulfill the first requirements, one can have the idea to move along the unit circle thetransition point between the swing and stance phase from an angle ρ ≥ 0 (see figure 3.5). Therefore,the angular rate of the hip at touch down will not be null anymore, but will be sin(ρ). Finally inorder to fulfill the last requirements, a transformation of the plan, in polar coordinates could beproposed :
R+× [−π,π]→ R+× [−π,π]�
rθ
��→
�r
fρ(θ)
� (3.19)
provided that fρ is a bijection of [−π,π] and that fρ is continuous, the unit circle will be left
30
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
unchanged. The following properties for fρ are also required :
fρ(0) = 0 (3.20)fρ(−π) = ρ −π (3.21)
fρ(π) = ρ +π (3.22)f �ρ(0) = 1 (3.23)
f �ρ(−π) = f �ρ(π) (3.24)
• Equation (3.20) avoids any displacement of the stance to swing transition point.
• Equations (3.21) and (3.22) move the stance to swing transition point and are a necessarycondition, fρ to be a bijection.
• Equations (3.23) and (3.24) insure a smoothness of the solution
One can found a unique 4th order polynom that satisfies the conditions (3.20-3.24) :
fρ(θ) =−ρ.
�θπ
�4+2.ρ.
�θπ
�2+θ (3.25)
The result of one of these transformations is seen in figure 3.6.
(a) Original trajectories (b) Images of the trajectories (c) Transform of a uniform mesh
Figure 3.6 – Transformation of a few trajectories with the function fρ , for ρ = π4 .
One should also notice that all the conditions are not met, fρ to be a bijection. However ifρ ∈
�0, π
2�, this is always the case.
31
Here the choice of extending at maximum the leg at each gait cycle has been made. The height ofthe CoM can easily be controlled, by adding a new parameter νi, and so not extend the leg as itmaximum (ϕknee
i = 0. Once again, the following equations are introduce, to provide only smoothchanges of the two level parameters :
λi = kλ
�λi−λ des
i
�(3.17)
νi = kν�
νi−νdesi
�(3.18)
3.1.2 Addition of the leg retraction principle
The next goal is to implement the leg retraction principle into the CPG. More formally, wewant that at touchdown, the gesture of retraction has already begun, i.e. : that both x < 1 andx > 0. Ideally we also want to parameterize this amount of retraction, by for example specifyingx(ttouchdown). In [SGH03], this parameter is the constant rotational speed of the leg.
Figure 3.5 – Implementation of the leg retraction principle. The swing to stance transition point should bemoved of an angle ρ ≥ 0.
However, implementing the leg retraction principle this way is rather difficult. In [SGH03]the angle (to the vertical) of the leg at APEX is precisely controlled, and the rotational speed isconstant. The first predicate is difficult to implement in a robot, and the second one isn’t reachablewithin the servo limits. Cheetah is intended to be really fast, at the limit of the servo specification.So it couldn’t be expected, the servo not to have a delay to reach some specific speed.
Further more, the convergence properties of this CPG are rather complicated, and defining newlimit cycle would be difficult. Then it may be simpler, in a first test, just to transform the output ofthe oscillators. Then the new behavior won’t interfere with the properties of the dynamical system.The trajectories generated by the new output should also be changed, as little as possible, from theoriginal one. Therefore the limit cycle (unit circle) must not be changed.
In order to fulfill the first requirements, one can have the idea to move along the unit circle thetransition point between the swing and stance phase from an angle ρ ≥ 0 (see figure 3.5). Therefore,the angular rate of the hip at touch down will not be null anymore, but will be sin(ρ). Finally inorder to fulfill the last requirements, a transformation of the plan, in polar coordinates could beproposed :
R+× [−π,π]→ R+× [−π,π]�
rθ
��→
�r
fρ(θ)
� (3.19)
provided that fρ is a bijection of [−π,π] and that fρ is continuous, the unit circle will be left
30
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Original trajectories Tansformed trajectories Tansformed mesh
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Optimization Definiton
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Optimization Definiton
• Using Particle Swarm Optimization
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Optimization Definiton
• Using Particle Swarm Optimization
• Differentiate between hindlimb and forelimb parameter
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Optimization Definiton
• Using Particle Swarm Optimization
• Differentiate between hindlimb and forelimb parameter
• Takes highest boundary as possible
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Name Applies to BoundariesHip::duty (d1,d2,d3,d4) [0.05,0.95]
Fore::Hip::amplitude (a1,a2) [0.0,2.0]Fore::Hip::offset (h1,h2) [−1.3,0.9]
Fore::Hip::refraction (ρ1,ρ2)�0.0, π
2�
Hind::Hip::amplitude (a3,a4) [0.0,2.0]Hind::Hip::offset (h3,h4) [−1.3,0.9]
Hind::Hip::refraction (ρ3,ρ4)�0.0, π
2�
Fore::Knee::amplitude (λ1,λ2)�0.0, π
2�
Fore::Knee::ThExtension (Θextension1 ,Θextension
2 )�π
2 , 3π2
�
Fore::Knee::ThContraction (Θcontraction1 ,Θcontraction
2 )�−π
2 , π2�
Hind::Knee::amplitude (λ3,λ4)�0.0, π
2�
Hind::Knee::ThExtension (Θextension3 ,Θextension
4 )�π
2 , 3π2
�
Hind::Knee::ThContraction (Θcontraction3 ,Θcontraction
4 )�−π
2 , π2�
Table 3.1 – Open parameters of the optimization
34
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
First Learned gait
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
First Learned gait
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• “Breaked gait” : lose all it speed at touchdown
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• “Breaked gait” : lose all it speed at touchdown• Forelimb are propulsive, hindlimbs lake foot clearance
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• “Breaked gait” : lose all it speed at touchdown• Forelimb are propulsive, hindlimbs lake foot clearance• Energetically unefficient : effector are taking energy of the
system
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Statistical method to find less influential parameters
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Statistical method to find less influential parameters
• Using PCA on the set of best particles over all iteration
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Statistical method to find less influential parameters
• Using PCA on the set of best particles over all iteration
• Principal Components will tell us if there is or not such parameters
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Statistical method to find less influential parameters
• Using PCA on the set of best particles over all iteration
• Principal Components will tell us if there is or not such parameters
• Caution : we may extracting the sctochastic characteristic of PSO !
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Statistical method to find less influential parameters
• Using PCA on the set of best particles over all iteration
• Principal Components will tell us if there is or not such parameters
• Caution : we may extracting the sctochastic characteristic of PSO !
• For the gait, the less influential is the extension threshold
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Insight for further improvement
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Insight for further improvement
• Move the CoM from the hind to the front
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Insight for further improvement
• Move the CoM from the hind to the front
• Increase stiffness and length of hindlimbs
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Insight for further improvement
• Move the CoM from the hind to the front
• Increase stiffness and length of hindlimbs
• Add sensory feedback to ensure “clean” stance phase
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Learning Trot Gait with Sensory
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Learning Trot Gait with Sensory
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Learning Trot Gait with Sensory
•Added Sensory Feedback
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Learning Trot Gait with Sensory
•Added Sensory Feedback•Optimized frequency
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Learning Trot Gait with Sensory
•Added Sensory Feedback•Optimized frequency•High instability of the gait in parameter space
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• CoM height is almost constant
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• CoM height is almost constant• Forelimb less propulsive, “almost” clean stance phase
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• CoM height is almost constant• Forelimb less propulsive, “almost” clean stance phase
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Dynamical analysis
• CoM height is almost constant• Forelimb less propulsive, “almost” clean stance phase• Energetically efficient for forelimb at least.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Further Work
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Further Work
• Test the gait to be stable for different duty ratio/ frequency.
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Further Work
• Test the gait to be stable for different duty ratio/ frequency.
• Mesure efficiency of the leg retraction principle by using stability criteria (APEX return map ....)
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Further Work
• Test the gait to be stable for different duty ratio/ frequency.
• Mesure efficiency of the leg retraction principle by using stability criteria (APEX return map ....)
• Add new actuator, like a spinal coord, scapula joint ...
Introduction Self-stabilization Strategy Model Update Control Design Optimization
Alexandre Tuleu Improvement of the Cheetah Locomotion Control
Introduction Locomotion Behavior in Nature Model Update Control Design Optimization
Questions ?