5. Modular robots
• reconfigurable
• swarm <> organism
• adapt to specific task
• production cost reduced
• self repair / replace
• more difficult to control
• lower performance on
specific tasks
advantages disadvantages
6. Modular robots
• Space / deep sea exploration
• Construction (larger architectural systems)
• Box of stuff
• Search and rescue in unstructured environments
applications
7.
8.
9. Trying to make a
neuroevolved modular robot
step over obstacles
13. Why stepping over
obstacles specifically?
• specific to walking behavior in legged robots
• walking behavior for legged robots is used
frequently
• small literature study/survey > not much ’abstract
motor control’ (turning, reversing direction,
ascending, descending and obstacle evasion).
14. Important terms
• Central Pattern Generators (CPGs)
• HyperNEAT
• based on NEAT
• compositional pattern producing network (CPPN)
• substrate
• hypercube: x1
* y1
* x2
* y2
15. Previous research
• HyperNEAT and extensions
• Walking behavior in quadrupeds, hexapods and
octopods
• Locomotion using Central Pattern Generators
• Manually created complex neural networks
• Staged or incremental learning
• HyperNEAT for Locomotion Control in
Modular Robots (2010)
20. Research question
Is it possible to neuroevolve modular robot
organisms that negotiate obstacles and step over
them (without using incremental learning?)
• How can we test if a robotic organism negotiates
obstacles instead of accidentally stepping over them?
• What should a fitness function for the wanted behavior of
stepping over obstacles look like?
21. Experiment setup
• experiment from HyperNEAT for Locomotion Control in
Modular Robots (2010) was taken and modified
• modular robot of 14 modules
• corridor with obstacles
changes:
• Rings of obstacles instead of corridor with obstacles
• obstacles are randomized before each simulation
• Webots collision detection plugin
• Fitness function modified
• Variants
22.
23.
24.
25. Experiment setup
Arena experiments
• obstacles surround the robot creature in such
a way that moving around them is impossible.
• 10 repeats of each variant
• 150 generations, 10 individuals
Arena ‘evasion’ experiments
• rings have more space between them
• rings have less obstacles
• 6 repeats of each variant
• 150 generations, 10 individuals
30. Stepping over
obstacles
How does one define 'stepping over' ?
• helper: raise legs higher
• very complex award system:
• raise leg near obstacle
• move over same obstacle
• lower leg after obstacle
what is near ?
what if part of obstacle ?
where is ‘after’ ?
51. Results conclusion
• baseline / no inputs perform nearly the same
• but baseline MBF performs better with less obstacles
• freeze output and larger substrate perform worse then
baseline
• substrate variation does not cover a lot of distance
53. Discussion
• Interpretation of results
• Answering the research questions
• Critique experiment design
• Suggestions for future work
• Things I’ve learned
54. Interpretation of results
• no sense of direction (circular movement)
• back away from obstacles
• especially in the case of larger substrate
variant
55. Research question
• How can we test if a robotic organism negotiates
obstacles instead of accidentally stepping over them?
• comparison in visual analysis: neuroevolution = black
box
• What should an objective fitness function for the wanted
behavior of stepping over obstacles look like?
• no objective fitness function at all !
56. Research question
Is it possible to neuroevolve modular robot
organisms that negotiate obstacles and step over
them (without using incremental learning?)
• yes, but:
• ‘performance’ is very low
• objective fitness function -> MOEA
• no objective fitness function at all.
57. Experiment design
critique
• autonomous module control (multiple brains)
• could still be modular
• no information distribution / communication
• less computation
• problem is a MOP: MOEA?
• objectivity issue stays.
60. Suggestions for future
work
• address experiment setup critique
• evolutionary benefit to stepping over
obstacles
• simple maze with obstacle shortcuts
• to resources (resource gathering)
• end of maze (freedom)
68. Novelty search
from http://eplex.cs.ucf.edu/noveltysearch/userspage/
• 'A deceptive problem is one in which it is hard to craft an
effective fitness function.’
• 'Novelty search is best suited for deceptive problems'
• great talk by Ken Stanley: http://www.santaferadiocafe.org/
science/2015/03/24/ken-stanley/
69. Things I’ve learned
• Time spent on implementation
• Spend time thinking abt. workflow: multiple experiments,
data extraction + analyzing.
• Simulation time
• lower ‘iteration’ speed
• triple check all settings before starting experiments
• Webots
• backward compatibility issues
• crashes
• collision bug
• Tooling / Frameworks