1. On the dynamics of human locomotion and co-design of lower limb assistive devices
Jesse van den Kieboom
Biorobotics Laboratory, EPFL, Lausanne
Supervisor: prof. Auke Jan Ijspeert
PhD oral exam, May 14, 2014

2. INtroduction
Source: youtube.com

3. introduction
Source: youtube.com

4. Introduction Motivation Body versus mind, embodied intelligence
Morphology determines how interaction takes place
Strong notion in natural system
Adaptation of morphology is a product of natural evolution

5. Introduction Motivation
EVRYON: evolving morphologies for human-robot symbiotic interaction (7th framework programme ICT Embodied intelligence)
Incorporate morphological design prominently into the design phase
Not always obvious, since performance is a combination of morphology and control
Take inspiration from natural processes to co-design morphology and control using dynamical simulations

6. Introduction
Topics
Development of a methodology for the co-design of bipedal machines, with a case study in wearable lower limb devices
Study of principles of human gait optimization and control
Rigorous modeling of coupled dynamical systems and rigid body dynamics suitable for locomotion and co- design

7. Overview Introduction Dynamical systems Human gait optimization
Co-design methodology for wearable devices
Conclusion

8. Dynamical systems

9. Dynamical systems
Generally: a system which changes over time
Fundamentally multi-domain
Specifically interested in
Control dynamics
Rigid body dynamics
Neuro-mechanical dynamics

10. Dynamical systems codyn - coupled dynamical systems
A framework for modeling and integrating multi-domain, coupled dynamical systems
Motivations
Free/open
Expressive/modeling
Performance
Educational

11. Modeled after IJspeert et al. (2007)
Dynamical systems
Modeled after IJspeert et al. (2007)

12. Dynamical systems codyn - Rigid Body Dynamics
Articulated rigid body dynamics essential tool in robotics research (modeling, simulation, control)
Hard problem, a variety of simulators existing today
Accuracy vs. performance vs. modeling effort

13. Dynamical systems Free/open Accurate Fast Multi-domain Hard contacts
Multi-body contacts
Modeling
Embedded
Bullet/ode + - ++
OpenSim
+/-
Robotran
Codyn

14. Dynamical systems codyn - Rigid Body Dynamics
Multi-domain: rigid body dynamics formulation indifferent from other dynamics, well suited for neuro-mechanical/musculo- skeletal modeling (Taga, 1995; Geyer, 2010)
Equations of motion derived from (Featherstone, 2008), transformed into coupled dynamics formulation of codyn
General purpose, 3D articulated rigid body dynamics
Entirely user extensible, customized joint models, contact models
State of the art: Availability of inverse, forward, closed chain, hard contacts, Jacobians, etc.

15. Dynamical systems

16. Dynamical systems

17. Dynamical systems

18. Dynamical systems

19. dynamical systems codyn - Performance
Generalized coordinates for RBD provides accuracy
When implemented naively, has very poor performance
Automatically translate models to efficient representation
Fast, optimized code
Suitable for Real Time systems
Suitable for low-resource, embedded systems (for example micro-controllers (Knüsel, 2013))

20. dynamical systems
codyn - Performance

21. dynamical systems

22. Human gait optimization

23. human gait optimization
What is the minimal, sufficient model for human gait optimization?
What are the objectives leading to stable, human gait?
What is the role of impedance modulation for human gait?
How prominent is the morphology for optimization of human gait?

24. human gait optimization
Extensively studied
Neuro muscolu-skeletal control using oscillators and entrainment (Taga, 1995)
Musculo-skeletal reflex models (Geyer, 2010)
Optimal control approaches (Mombaur, 2010)
Passive dynamic walkers (McGeer, 1990; Wisse 2005; Kuo 2007)

25. human gait optimization
Motivation
To obtain a minimal, sufficient optimization model to explain human locomotion
Using simple, local control only, few parameters
Main inspiration: passive dynamic walkers
Collins (2001)

26. human gait optimization
Hypotheses
Human like gaits occur implicitly using impedance control by optimizing only for mechanical energy expenditure
Gait quality increases with increasing modulation of impedance
Energy expenditure decreases with increasing modulation of impedance

27. human gait optimization
Level of abstraction
2D, planar walking
Segmented legs (upper leg, lower leg, foot)
Perfect torque actuators at the joints

28. human gait optimization
Control
Joint level variable impedance control
Optimize reference trajectory, variable stiffness and variable damping
Various levels of variable impedance: const, step, ppoly

29. human gait optimization
Particle Swarm Optimization (Kennedy, 1995)
Codyn for modeling of control and rigid body dynamics
Objective
Until 1
time
max time 2
speed match
95% 3
std speed
0.1 4
assist time 5
non contact time 6
torque -

30. human gait optimization

31. human gait optimization
Single step (stance to stance)

32. human gait optimization
Hip, knee, ankle
Single step (stance to stance)

33. human gait optimization
Hypotheses
Human like gaits occur implicitly using impedance control by optimizing only for mechanical energy expenditure
Gait quality increases with increasing modulation of impedance
Energy expenditure decreases with increasing modulation of impedance ✓ ✓ ✓

34. human gait optimization
Transfer of methodology to a human-like platform
Effect of morphological differences to bipedal gait obtained from optimization
Role of variable impedance under perturbation

35. human gait optimization
Transfer of methodology to a human-like platform
Source: IIT website

36. human gait optimization
Transfer of methodology to a human-like platform
Same optimization methodology except:
Improved objectives due to hard-contact modeling in codyn (remove std speed and non-contact-time)
Optimize for (mechanical) cost of transport instead of desired speed
Adaptation of foot length to size of the robot
Limit maximum torques to platform limits (30N)

37. human gait optimization

38. human gait optimization
Observations
Foot length important for obtaining robust gaits
Mass distribution influences stability, making it harder to optimize
Resulting gait slow, maximum torque limits

39. human gait optimization
Transfer of methodology to a human-like platform

40. human gait optimization
Transfer of methodology to a human-like platform
Role of variable impedance under perturbation
Apply force perturbations during window in swing phase
N(80, 5) Newton
N(150, 50) Milliseconds
Same optimization methodology

41. human gait optimization
Transfer of methodology to a human-like platform

42. human gait optimization
Hip, knee, ankle

43. human gait optimization
Transfer of methodology to a human-like platform
Successfully optimized human-like gait to robot morphology
Foot length is very important for obtaining gait
Mass distribution influences stability
Variable impedance modulation significant under perturbation

44. Morphology/control co-design

45. co-design
Motivation
Many successful (some commercial) wearable lower limb devices
Impressive results: allows for paraplegics to walk again
Design is always anthropomorphic
Potential for exploitation of morphology
Ekso
ReWalk
Rex

46. co-design Problem domain
Development of a co-design methodology, using evolutionary inspired processes
EVRYON: use case for a non-anthropomorphic, lower limb wearable device for locomotion
A framework for the iterative co-design of wearable devices, focussing on open-ended exploration of solutions

47. co-design

48. co-design

49. co-design Level of abstraction
Same human model as in previous studies (2D, planar)
Looking at mechanical dynamics only
Augmentation of the hip and knee only (2 dof parallel structures)
Parallel structures involving 3 human body attachment joints, 3 free wearable robot joints and 4 wearable robot segments

50. co-design

51. co-design Morphological space Wearable robot attachment placement
Active joint placement
Wearable robot segment lengths
Wearable robot initial joint angles

52. co-design

53. co-design
Solutions obtained were never self-stable, torso was constraint to be upright (mass-distribution)
Ground contact modeling problematic, unstable
Almost all solutions made use of structural singularities
Main conclusion: promising, however available tools are insufficient

54. co-design
Optimization no longer only considering continuous parameters only
Particle Swarm Optimization in itself not suitable
Consider Genetic Algorithms (Goldberg, 1987) or Genetic Programming (Koza, 1992)
Many parameters which significantly alter performance, PSO found to perform better for locomotion (Bourquin, 2004)
A novel PSO for co-design

55. co-design Meta morphic particle swarm optimization
Extension of PSO principles to simultaneous exploration/optimization of solution structures and its parameters
Suitable for problems where solution structures can be enumerated

56. co-design Structural parameters (MMPSO) Morphological parameters
Topology
Actuator placement
Morphological parameters
Attachment location
Wearable robot joint location
Control parameters
Joint angle pattern, stiffness pattern, damping pattern

57. co-design

58. co-design
Topo
Mass
CoM (up)
Segment size
Power 2
83.3 (+13.3)
1.07 (-0.03)
462 5
83.9 (+13.9)
1.02 (-0.08)
748 3
83.5 (+13.5)
1.03 (-0.07)
1018

59. co-design
Development of a complete framework for co-design (dynamics/control, optimization, simulation)
Successfully optimize human like gaits with parallel structures
Automatic and simultaneous exploration of solution structures and their parameters
Mass distribution particularly important
Actuator placement favors actuator on the torso

60. Conclusion

61. conclusion Main contributions
Open and freely available, framework for modeling of multi- domain coupled dynamics systems
State of the art, competitive rigid body dynamics simulator
Novel particle swarm optimization based algorithm for co- optimization of solution structures and their parameters
Robust optimization of human gait from first principles using simple, local impedance control
Front to end framework for the co-design of morphology and control of robotic structures

62. conclusion Discussion Human locomotion Extension of methodology to 3D
Transfer of control to robotic platform
Feedback based variable impedance control

63. conclusion Discussion Co-design Methodology shown to be feasible
Should provide input in a larger, iterative design process
Open-ended search does not provide complete solutions

64. conclusion Discussion Co-design Methodology shown to be feasible
Should provide input in a larger, iterative design process
Open-ended search does not provide complete solutions
Only looked at mechanics so far, no human-in-the-loop
Integration of neuro/musculo skeletal models, studying human adaptation

65. conclusion Lessons learned
Within constraints of EVRYON, obtained wearable robot solutions did not provide the aimed for dynamic symbiosis purely using evolutionary algorithms
Applying the methodology to smaller subsystems may be more promising
Use an iterative design approach (engineer-in-the-loop), rather than a front-to-end automatic approach
Importance of human-in-the-loop cannot be ignored, but tools need to be suitable for such simulations

66. Questions

67. Co-design

68. References
Ijspeert, A. & Crespi, A. & Ryczko, D. & Cabelguen, J.-M. (2007). From swimming to walking with a salamander robot driven by a spinal cord model. Science,315, 1416—1420.
Taga, G. (1995). A model of the neuro-musculo-skeletal system for human locomotion. I. Emergence of basic gait.Biological Cybernetics,73, 97—111.
Geyer, H. & Herr, H. (2010). A Muscle-Reflex Model That Encodes Principles of Legged Mechanics Produces Human Walking Dynamics and Muscle Activities. Neural Systems and Rehabilitation Engineering, IEEE Transactions on,18, 263—273.
Mcgeer, T. (1990). Passive dynamic walking. the international journal of robotics research,9, 62—82.
Wisse, M. (2005). Three additions to passive dynamic walking: actuation, an upper body, and 3D stability.International Journal of Humanoid Robotics,2, 459—478.
Kuo, A. D. (2007). The six determinants of gait and the inverted pendulum analogy: A dynamic walking perspective.Human movement science,26, 617—656.
Mombaur, K. & Truong, A. & Laumond, J. (2010). From human to humanoid locomotion—an inverse optimal control approach. Autonomous robots,28, 369—383.
Knüsel, J. (2013). Modeling a diversity of salamander motor behaviors with coupled abstract oscillators and a robot. EPFL.
Goldberg, D. E. & others , . (1989). Genetic algorithms in search, optimization, and machine learning. Addison- wesley Reading Menlo Park,412.
Koza, J. R. (1992). Genetic programming: on the programming of computers by means of natural selection. MIT press,1.
Bourquin, Y. & Ijspeert, A. J. & Harvey, I. (2004). Self-organization of locomotion in modular robots. Unpublished Diploma Thesis, epfl. ch/page html.
Kennedy, J. & Eberhart, R. (1995). Particle swarm optimization. Proceedings of IEEE international conference on neural networks,4, 1942—1948 vol.4.
Featherstone, R. (2008). Rigid body dynamics algorithms. Springer New York,49.

69. human gait optimization
Transfer of methodology to a human-like platform
Step controller kinematics
Single step (stance to stance)

70. HUMAN GAIT OPTIMIZATION