Colloque IMT -04/04/2019- L'IA au cœur des mutations industrielles - Contrôleurs neuronaux plastiques pour l'émergence de coordinations motrices dans l'interaction physique et sociale humain/robot
Colloque IMT -04/04/2019- L'IA au cœur des mutations industrielles - Session Robotique, Perception, Interaction: Contrôleurs neuronaux plastiques pour l'émergence de coordinations motrices dans l'interaction physique et sociale humain/robot. Présentation par Patrick Henaff (IMT Mines Nancy)
Double Revolving field theory-how the rotor develops torque
Colloque IMT -04/04/2019- L'IA au cœur des mutations industrielles - Contrôleurs neuronaux plastiques pour l'émergence de coordinations motrices dans l'interaction physique et sociale humain/robot
1. CONTRÔLEURS
NEURONAUX PLASTIQUES
POUR L'ÉMERGENCE DE
COORDINATIONS
MOTRICES DANS
L'INTERACTION PHYSIQUE
ET SOCIALE
HUMAIN/ROBOT
Patrick Hénaff
Mines Nancy
LORIA
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 1
2. 1. INTRODUCTION
• Human rhythmic movements
• Interpersonal coordination
• Synchrony and synchronization
2. RHYTHMIC MOVEMENTS OF ROBOTS
• Central pattern generators (CPG)
• Computational model of CPG
3. LEARNING MOTOR COORDINATION
• Implementation of the bio-inspired controller
• Experiments:
• Waving back with a robot
• Coordination of complex rhythmic movements
• Discrete movements
4. CONCLUSION & FUTURE WORKS
OUTLINE
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 2
4. HUMAN PHYSICAL CO-MANIPULATION FOR
RYHTHMIC TASKS
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 4
Italian Institute of Technology, Dep of Advanced robotics, 2016
• Rhythmics tasks are often base on regular motions without
need of accurate trajectories
• Control of antropomorphic robots can be inspired from
human motor nervous system?
5. HUMAN MOVEMENTS
• Discrete movements :
• Static (taking object)
• Dynamic (sport, basket)
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
• Rhythmic movements are very primitive and automatic
• Interactions and interpersonal coordinations are often
based on rhythmic movements
Rhythmic movements and discret movements
• Rhythmic movements :
• Locomotion(walking, running)
• Cleaning, brushing
• Sports (basket)
Rhythmic movements Discret movements
Upper limbs « non regular » « regular »
Lower limbs « regular » « non regular »
5
6. MAIN PROPERTIES IN INTERPERSONAL INTERACTION
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
Involuntary interpersonal coordination
Compromise between engine control and interaction
2
1
An interaction is always based on two main aspects:
• The individual part of each participant
• The share of the coupling in the pair of individuals
Phenomenon of emergence of coordination is so powerful that humans can not
avoid involuntary dyadic motor coordination
interpersonal coordination is based on
synchronization of movements
6
Walk Walk applause
7. WHAT IS SYNCHRONIZATION OF MOVEMENTS?
Interpersonal and Interlimb Coordination
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 7
extracted from [Richardson et al, 2007]
Two spontaneously
stable coordination
patterns appear:
• in-phase (0°)
• anti-phase (180°)
extracted from [Dumas et al, 2014]
Interpersonal synchronization acts like a dynamical system
based on 2 coupled non-linear oscillators
9. EXAMPLE OF INTERPERSONAL SYNCHRONIZATION IN
PHYSICAL INTERACTION : THE HANDSHAKING
Phase 1 – SoH
Start of Handshake
Phase 2 – PhC
Physical Contact
Phase 3 – MS
Mutual Synchrony
Phase 4 – EoH
End od Handshake
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 9
Average of 5 handshakes.
Subject I Subject II
• Synchronization exists in interpersonal rhythmic
physical interactions
• Handshaking acts as coupled nonlinear oscillators
[ Melnyk and Henaff, 2014, 2019]
[ Tagnes and Henaff, 2016]
10. [Zehr, 2006]
Rhythmic or discrete
movements?
[McCrea & Rybak, 2008]
HUMAN RHYTHMIC MOVEMENTS ARE CONTROLLED
BY CENTRAL PATTERN GENERATORS (CPG)
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 10
CPG for lower limbs
(locomotion)
CPG for upper limbs
CPG for upper limbs ?
Bilateral left-right intercations
[Rybak et al. 2015]
?
12. Rhythmic level
Formation level
motor level
COMPUTATIONAL MODEL OF CPG ?
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
CPG for lower limbsCPG for upper limbs
Computational CPG for
one joint
12
[Nassour, Hénaff 2014]
• Model of oscillating neuron?
• Model of sensory feedback?
13. MODEL OF RHYTHMIC NEURON :
13AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
Neuron model = Non-linear oscillator (Rowat & Selverston model, 1993)
σf
σs
Marder et al. (2001)
intrinsic
properties of
biological
neurons
14. Input signal
CPG MODEL FOR ONE JOINT
Rhythmic generator cells : Rowat-Selverston model (Van der Pol formalism):
Interneurons :
Hebbian Plasticity :
Frequency learning based on Righetti’s rule:
Amplitude learning:
Input gain learning:
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
14
[Jouaiti, M., Caron, L., and Henaff, P. (2018).
16. WAVING BACK EXAMPLE
• Human waves at robot
• Optical flow of the hand detected
• Robot waves back and adapts to the human frequency
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
16
COUCOU !
HELLO!!!
17. WAVING BACK : CONTROL ARCHITECTURE
2 joints controlled :
CPG1 input (shoulder) : optical flow
CPG2 input (elbow): output of CPG1
CPG output: joint angular position
interest of plasticity:
faster synchronization
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 17
[Jouaiti and Hénaff, 2018b]
shoulder
elbow
18. WAVING BACK EXPERIMENT
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
18
Synchronization between joints and optical flow Hebbian plasticity
19. AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
19
WAVING BACK EXPERIMENT
plasticity influences learning speed
with amplitude and synaptic weight learning.
without amplitude and synaptic weight learning.
Phase portrait of rhythmic cells outputs
shoulder,
shoulder,
elbow
elbow
Synchronization Measures (10 Wavings)
Phase lock value
20. COORDINATED COMPLEX MOVEMENTS
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 20
shoulder
elbow
shoulder
Circles at different speeds
“infinite” sign
21. Can CPGs achieve both rhythmic
and discrete movements ?
21AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
22. OSCILLATING NEURON CAN BEHAVE AS A PID
CONTROLLER
22AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
Non oscillating :
23. EXPERIMENT
23
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
Joint
phase
portrait
CPG
phase
portrait
Robotic arm coupled to a camera detects a target and raises towards it
when the target is reached, the arm oscillates at its own intrinsic frequency
24. RESULTS
CPG can transition smoothly between both states by
changing a single parameter
24
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
Phase portrait of
the output of the
CPG (extensor
and flexor) and of
the joint position
Joint phase portraitCPG ouputs phase portrait
25. CONCLUSION & FUTURE WORKS
versatility and reliability of CPG controller that can adapt to varying
frequencies
dynamic control without any dynamic model of the robot
can achieve both discrete and rhythmic movements
can also act as position or velocity controller
emergence of synchronisation (motor coordination) intrapersonnal and
interpersonnal
Approach can be used in human/robot physical cooperation for several
different rhythmic tasks in dangerous environment:
Cleaning, Scrubbing, Brushing
AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY 25
Dynamic control based on CPGs is very efficient for robot
rhythmic motions :
• for physical interactions with human and environment
• but without need of accurate trajectories
26. AVRIL 19 - IMT - P. HÉNAFF - MINES NANCY
Thanks for your attention…
26