Locomotion control for a quadruped robot based on motor primitives

1
Locomotion control for a quadruped robot based on
motor primitives

• Verena Hamburger

2
Advantages of motor primitives

• Biologically inspired (mankind, frog etc.)
• Evasion of online trajectory planning
• Simple solution of Bernstein-Problem
• Facilitates the generation and perception of
  movements (-gt mirror neurons)
• Robust classification of faulty movement pattern
  (even prediction in infants)
• Superposition comparable to linear approximation
  and Fourier analysis

3
Importance of morphology

• Behavioural diversity
• Cheap design
• Ecological balance
• Emergence of behaviours

4
Aim

• A methodology together with morphology and task
  independent measures which aims at the
  development of quality criteria and design
  guidelines for motor primitives and fitting
  morphologies

5
Proceeding

• Preprogrammed experiments
• Generation of vocabularies
• Selection of morphologies
• Behavioural diversity
• Robustness against variation of morphology,
  position and inclines of 22.5 and 30
• Learning progress of Reinforcement Learning
• Development of morphology and task independent
  measures

6
Preprogrammed experiments toexplore the
potential of the given hardware
7
MiniDog6M

• 6 servo motors shoulder, hind, spine
• Spring approximates artificial muscle
• Control cycle 1 sec because of self-stabilisation
• Identical leg design one active and one passive
  dof
• Acceleration sensor (static and dynamic) in the
  head

8
(No Transcript)
9
Generation of vocabularies
10
Flexibility-Index Flx

• Some primitives do not specify all motors, but
  uses stop symbol/Dont-Care x instead
• gt Flexibility-Index

11
Transfer into simulation Selection of
morphologies
12
Investigated morphologies

• Head has biggest influence on standing up
• Original can lie on left, right, back and head
• Vertical cannot roll over on back and not stand
  on head
• Round can roll over on the back most easily, but
  cannot lie on back nor stand on head
• gt different numbers of possible initial
  positions
• gt different effort for learning

13
Behavioural diversity Robustness against
variation of shape and posture
14
Behavioural diversity

• No uniform definition for intelligence, but
  behavioural diversity is part of all current
  definitions
• Increases adaptability
• Not only amount is important, but also a balance
  between heterogenity and homogenity of solutions
• Mean duration of solutions as non-functional
  criteria

15
Behavioural-Diversity-Index BDI

• BDI is the product of the total amount of legal
  sequences with the mean diversity factor D
• D is a measure for the heterogenity of sequences.

16
Results (1)

• Flx gt 0 results in higher BDI than Flx 0
• Compatibility of sequences heavily depends on
  vocabulary (Sometimes even symmetrical
  tendencies)
• Position stable a sequence is successful for
  more than one initial position of the same head
• Form stable a sequence is successful for more
  than one heads of the same initial position

17
Results (2)

• Only vertical head supports position stability
  for all vocabularies
• Form OR position stable sequences involve
  primitives with and without Don't-Care (Flx 0)
• Form AND position stable sequences involve only
  primitives without Don't-Care (Flx 0)

18
Robustness on inclines of 22.5 and 30
19
Results

• Weight distribution and centre of mass are
  identical in all three heads
• Original and the round head perform equally or
  better than on flat terrain gt most robust
• Vertical head solutions for Left and Right are
  exchanged (sometimes one side behaves as if with
  original head)
• gt Morphological effect which is purely grounded
  in the shape

20
Learning progress Q-Function is
linearly approximated on the basis of a
RadialBasisFunction-Network
21
Results (1)

• BDI gt 10 guarantees success within max. 100 of
  200 given episodes
• Vocabularies with low BDI often do not find a
  solution for one or more initial positions

22
Results (2)

• Flx gt 0 outweighs the higher learning effort
  through more initial positions
• Flx 0 Flx gt 0

23
Conclusion

• Only few vocabularies are equally good at all
  features and morphologies
• Use of Don't-Care (Flx gt 0) is good for BDI and
  learning progress, but not for robustness
• Effects of shape and material must already be
  considered in the design phase
• No need for adaptation in case of proper design
  and enough behavioural diversity
• Currently cheapest design and control of a
  running quadruped

24
To Do

• Application of this methodology to many different
  tasks, morphologies, learning methods and
  vocabularies
• Examination of robustness on different ground
  properties e.g. elastic, irregular
• Additional investigations of sensory-motor-coordin
  ation

25
The end.