Bio-inspired control and sensorsweb.eng.ucsd.edu/~ngravish/MAE207/PDF/Lectures/10_web.pdf · •...
Transcript of Bio-inspired control and sensorsweb.eng.ucsd.edu/~ngravish/MAE207/PDF/Lectures/10_web.pdf · •...
Bio-inspired control and sensors
5/7/2019
THE STONE AGES
WHAT ARE CONTROL SYSTEMS ?
N. J. Cowan, M. M. Ankarali, J. P. Dyhr, M. S. Madhav, E. Roth, S. Sefati, S. Sponberg, S. a. Stamper, E. S. Fortune, and T. L. Daniel, “Feedback control as a framework for understanding tradeoffs in biology,” Integr. Comp. Biol., vol. 54, pp. 223–237, 2014.
WHAT ARE CONTROL SYSTEMS ?
CONTROLALOGRITHM
SENSORS
BODYDYNAMICS
INPUT OUTPUT
+
-
MOTIVATION
Robotic control could be better
Better control would mean better robots
MOTIVATION
Study animals to develop better controllers
MOTIVATION
• How are animals achieving control?
• How are robots controlled?
(These questions are very broad)
BIOLOGICAL CONTROL HARDWARE
[Wikipedia]
[Santiago Ramón y Cajal (from wiki)]
Purkinje cell (Cerebellum)
Interneuron (Spine Gray Matter)
[Pearson]
Sensory Neuron
[Wikipedia]
BIOLOGICAL CONTROL HARDWARE
Biological sensory systems are slow
Mass (kg)
LOW LEVEL CONTROL: BIOLOGY
• Nerves have non-negligible transmission delays
• Speed:• 0.6 m/s (pain signals)• 100 m/s (large diameter motor axons)
• Not even counting synapse delay
Nerves are slow!3 million times slower than wire
5.8 ms for neural impulse to travel from spine to hand muscles
0.5 ms required for accuracy [Hore, Watts & Tweed ‘96]
Nerve impulse at spine
FEEDFORWARD CONTROL
Activation onsetbefore impact
[Organization of Movement, Krakauer & Ghez 2000]
HIERARCHAL CONTROL
• Some control needs to happen fast
• Some control needs lots of information
• You can’t really have both ☹
• Solution: Hierarchal Controllers
E. Todorov, W. Li, and X. Pan, “From task parameters to motor synergies: A hierarchical framework for approximately optimal control of redundant manipulators,” J. Robot. Syst., vol. 22, no. 11, pp. 691–710, 2005.
HIERARCHAL CONTROLROBOT ANIMAL
HIGH
MID
LOW
Force/Position
Pose target
GRASP TASK
DYNAMIC PLANNER
MOTOR CONTROLLER
MOTOR CONTROLLER …
????
???
GRASP TASK
CEREBELLUM(probably)
MOTOR POOLS
1000 Hz
100 Hz
25 Hz
HIERARCHAL CONTROL: BACKGROUND
• Two main approaches to control hierarchies:
• Centralized: simple, easy to modify, bandwidth limited
• Decentralized: offloads computation, can reduce delay
HIERARCHAL CONTROL MIT CHEETAH
Image courtesy of “High speed trot-running: Implementation of a hierarchical controller using proprioceptive impedance control on the MIT Cheetah” by Hyun et al.
• The MIT Cheetah uses a gait-pattern modulator to develop patterns for the leg trajectories, similar to a CPG
• The lower level controller is an impedance controller
• Many robots use a roughly similar structure but with different implementations details, e.g. trajectory arrays, set points, optimal control, PD control.
LOW-LEVEL CONTROL: MUSCLE STRETCH REFLEX• Involuntary muscle contraction to a
sudden stretch sensed by muscle spindles
• This is basically a load dependent response to input
• The purpose of this is mostly for postures and movements that are robust to disturbances
• Both excites and inhibits opposing muscle groups to incite motion
• Both this muscle reflex and the voluntary muscle movements controlled by the brain are responsible controlling motions
[Human Physiology: An integrated Approach, Silverthorn 2012]
LOW-LEVEL CONTROL: MUSCLE FLEXION REFLEX• Involuntary muscle
contraction due to a painful stimulus, such as stepping on a lego
• This muscle reflex is mostly for protection
• These reflexes can trigger reactions on the other limb to support the increased weight
• Reactions are quicker than it takes for the pain signal to reach the brain
[KOREAN MEDICAL LIBRARY ENGINE (HTTP://WWW.KMLE.CO.KR)]
LOW-LEVEL CONTROL: REFLEXES IN THE COCKROACH
• Cockroaches have a turn response for wind or sound stimulus• They tend to jump and turn away from the response, preparing to run• This behavior is still observed when the head is removed
[Hill 1989]
LOW-LEVEL CONTROL: DO ROBOTS HAVE REFLEXES?• Sort of
• Reflexes are mostly needed for speed and protection
• Robots typically don’t need to go faster, but could use a decentralized control architecture if they do
• An interrupt based protection scheme can be used here
LOW-LEVEL CONTROL: HyQ FOOT “REFLEX”
• It modifies its trajectory if it senses an obstacle
• Not really the same structure as an animal reflex
[Focci et al. CLAWAR 2013]
IMPEDANCE CONTROL: BACKGROUND
• Key idea: instead of focusing on perfectly or optimally tracking some trajectory, we can instead focus on HOW we interact with the environment
• The impedance control paradigm focuses on specifying the relationship between a flow input (position or velocity) and an effort (force or torque)
• This paradigm is useful for understanding both animal and robotic manipulation tasks
IMPEDANCE CONTROL IN HUMANS
• Hogan developed impedance control from a motivation for how humans interact with their environment
• Idea is that muscle stiffness is modulated to vary impedance
• Muscle stiffness is varied primarily through muscle co-contraction, posture selection, and stretch reflexes
• This approach generally assumes that the brain selects an impedance for the lower level controls to assume for a given task.
Image courtesy of “Impedance Control: An Approach to Manipulation” by Neville Hogan
IMPEDANCE CONTROL IN ROBOTS
• The MIT cheetah implements an impedance control algorithm for it’s legs
• The apparent impedance to the environment is varied depending on the task
• Impedance control fits into our distributed control architecture
• Downside: you have to choose an impedance. This is not always straight forward!
[Hyun et al. IJRR 2014]
IMPEDANCE CONTROL IN ROBOTS
• The MIT cheetah’s desired impedance is a coupled linear and radial spring-damper system
• The MIT cheetah receives leg trajectories from the higher level controller and
Images courtesy of “High speed trot-running: Implementation of a hierarchical controller using proprioceptive impedance control on the MIT Cheetah” by Hyun et al.
RYTHMIC CONTROL
Name as many rhythmic behaviors as you can. ! Is the behavior purposeful or involuntary?
! What enforces the periodicity? (Sensory feedback? Environmental interaction? Something else?)
! Alternatively, what causes the periodicity to break down?
RYTHMIC CONTROL
• Link behavior to a steady rhythm
• Be able to alter that rhythm if conditions vary
!Welcome to central pattern generators (CPGs)!
http://media.giphy.com/media/52LkAuJYMzjt6/giphy.gif
http://images.viralnova.com/000/074/910/blue-ribbon-eel.gif
https://media3.giphy.com/media/O1Mfoj3I5eAo0/200.gif
http://stream1.gifsoup.com/view1/1758417/chewing-cow-o.gif
http://stream1.gifsoup.com/view3/1979255/breathing-o.gif http://upload.wikimedia.org/wikipedia/
commons/6/6b/Elephant_walking.gifhttp://i.imgur.com/CAZcD03.gif
RYTHMIC CONTROL: CPG REGULATED BEHAVIORS
• Consist of distributed neural networks of multiple coupled oscillators
• Produce rhythmic behavior without the aid of:• Higher-level
sensory feedback• Higher-level control
[5] Rossignol et al, 2006
RYTHMIC CONTROL: CPG REGULATED BEHAVIORS
IB 222
RYTHMIC CONTROL: NEURONAL RYTHMICITY
Two (possibly) competing main ideas:• Pacemaker neurons:
individual oscillatory neurons
• Network pacemaker: synaptic connections in neural network generate rhythm• Half-center model: two
neurons reciprocally inhibit each other
5/11/2015
RYTHMIC CONTROL: CPGS SHAPED BY SENSORY FEEDBACK
• CPG provides rhythmic base for behavior
• Feedback keeps CPG and body in sync by:• Modulating phase &
frequency based on environmental changes
• Allowing locomotion mode to change as required/desired
http://birg2.epfl.ch/movies/SIMS/sal_s2w.htm (Ijspeert2007)
RYTHMIC CONTROL: BENEFITS OF CPGS
• Reduced time delay: signals only loop through spinal cord
• Smaller control signal dimensionality: CPG control is simpler than muscle control
• Reduced bandwidth required between higher control centers & CPG
RYTHMIC CONTROL: NEURAL BIOLOGICAL CPG MODELS
Three primary modelling paths:
• Connectionist: Use simple neuron models to explore rhythmogenesis
• Coupled oscillator: Explore impact of coupling type on dynamics of multiple oscillators
• Neuromechanical: Include model of body, allowing study of sensory feedback
[2] Ijspeert et al, 2007
RYTHMIC CONTROL: CPGS IN ROBOTICS
Do CPGs impart the same benefits to robotic control as they do to biological control?
!Generally yes, as well as some additional robotics-specific opportunities
[3] Ijspeert and Crespi, 2007
RYTHMIC CONTROL: ATTRACTIVENESS OF CPGS IN ROBOTICS
• Limit cycle behavior is very stable• Distributed implementation is good for modular robotics• Reduced number of control parameters• Easy sensory feedback integration• Ideal for exploring learning/optimization
[1] Ijspeert 2008http://biorobotics.ri.cmu.edu/projects/modsnake/gaits.html
RYTHMIC CONTROL: COMPONENTS OF CPGS
1. General architecture: # of oscillators/neurons, choice of position or torque control
2. Coupling type & topology
3. Waveforms: can be altered with filters
4. Effect of input signals
5. Effect of feedback
CPG exampleKuramoto oscillator model
d✓idt
= !i +K
N
NX
j=1
sin (✓j � ✓i) , i = 1 . . . N
CPG exampleModified Kuramoto oscillator for a legged system
✓̇i = 2⇡fi +6X
j=1
Kaij sin (✓j � ✓i � �ij)
Limb frequency
Limb coupling strength
Limb phase offset
RYTHMIC CONTROL: MULTILEGGED ROBOTS
WITHOUT FEEDBACK WITH FEEDBACK
[Tarapore2014]
[4] Kimura et al, 2007
RYTHMIC CONTROL: QUADRUPEDAL ROBOTS
RYTHMIC CONTROL: BIPEDAL ROBOTS
[https://sites.google.com/site/robodreamer/projects]
IB 222[3] Ijspeert and Crespi, 2007
RYTHMIC CONTROL: EEL ROBOTS
5/11/2015
[2] Ijspeert et al, 2007
RYTHMIC CONTROL: SALAMANDER ROBOTS
5/11/2015
IB 222[2] Ijspeert et al, 2007
RYTHMIC CONTROL: MORE SALAMANDER ROBOTS
5/11/2015