1) The document discusses an approach to the targeted environment-specific evolution of robot components using additive manufacturing and evolutionary algorithms.
2) A genetic algorithm is used to evolve robot leg designs optimized for different environments like soil, gravel, and fluid using a physics simulation.
3) The results show legs evolved in the soil environment perform best in soil, while legs evolved in other environments like gravel or fluid are better suited to those environments than a leg designed for a different terrain.
Artificial intelligence in the post-deep learning era
Towards the targeted environment-specific evolution of robot components
1. Towards the Targeted Environment-
Specific Evolution of Robot Components
DATA61
Jack Collins1,2
PhD Student
Dr. David Howard1
Research Scientist
Frederic Maire2
Senior Lecturer
Wade Geles1
Industrial Trainee
1 with Data61, Commonwealth Scientific and Industrial Research Organisation (CSIRO), Brisbane, Australia
2 with the Science and Engineering Faculty, Queensland University of Technology (QUT), Brisbane, Australia
2. • Design of robots typically involves:
• Expert knowledge
• Many man-hours
• Product iterations
• Influenced by human preconceptions
• i.e. This shape performs well in this environment because …
Motivation
2 |
4. • Evolutionary Robotics – “Automated creation of autonomous
robots inspired by Darwinian Evolution”
• Brain and Body Co-evolved
• Evolutionary Art
• Our Approach
– Mature Control Stack
– Targeted Evolution
– No Controller Evolution
4 |
Karl Sims, 1994
Jeff Clune, 2013
8. 1: High
• Multiple Beziers with varying control points and thicknesses.
Encoding (Representation)
Bezier graphic from http://graphics.cs.ucdavis.edu/~joy/GeometricModelingLectures/Unit-4/Unit4.html
8 |
9. 1: High
• Multiple Beziers with varying control points and thicknesses.
Encoding (Representation)
Bezier graphic from http://graphics.cs.ucdavis.edu/~joy/GeometricModelingLectures/Unit-4/Unit4.html
9 |
10. 1: High
• Multiple Beziers with varying control points and thicknesses.
Encoding (Representation)
Bezier graphic from http://graphics.cs.ucdavis.edu/~joy/GeometricModelingLectures/Unit-4/Unit4.html
10 |
11. 1: High
• Multiple Beziers with varying control points and thicknesses.
Encoding (Representation)
Bezier graphic from http://graphics.cs.ucdavis.edu/~joy/GeometricModelingLectures/Unit-4/Unit4.html
11 |
12. Fitness Function
• Fitness of an individual is calculated using:
– 𝜏: Accumulated Torque
– 𝛿: Percentage of occupied voxels
• A lower fitness is the objective
𝑓 =
𝜏
𝑛 𝑠𝑡𝑒𝑝𝑠
+
𝜏
𝑛 𝑠𝑡𝑒𝑝𝑠
×
𝛿
5
12 |
13. Genetic Algorithm
• Combination of the parents and children.
• Deletion of the worst performing half.
• Elitism retains fittest parent.
• Selection through tournament selection.
• Crossover using a two-point crossover method, with the two crossover
points randomly chosen values within the range of the shortest parents
genotype.
• Mutation
– Gaussian distribution of each Bezier control point
– Change Bezier thickness (20%)
– Add/remove control point (20%)
– Add/remove Bezier (10%)
13 |
15. Experimental Setup
• Population of 20 legs
• Optimised for 100 generations
• 30 experiments in total
• 10 Experiments per environment
– Soil - 7 Hours
– Gravel – 2.7 Days
– Fluid – 9.4 Days
• Parallelised using OpenMP
• Run on HPC
• Across 600 cores
• Not GPU enabled
15 |
-Typically the design of robots requires …
-And this is influenced by …
The wide spread adoption of Additive Manufacturing
Rapidly expanding assortment of construction materials
-Evolutionary Robotics is
-Early work by Jeff Clune
-Evolutionary Art …. PicBreeder … EndlessForms
-Our approach borrows from Evolutionary art in that we only evolve the morphology of robot components
-Targeted evolution of components that face “ENVIRONMENTAL PRESSURES“
-This allows us to utilise the mature control stack
-OBJ for Project Chrono
-smooth
-Decimate (10000 to less than 2000 faces)
-Transform and Rotate
-Two terms
-Torque
-Occupied Voxels
-Remove redundant material
-3D Printing
Over 100 Days into 9.4 days
-Soil legs being generally skinnier than the other two; as the leg cannot 'fall through' the soil, the second term in the fitness function has more influence on the final morphology.
-Gravel legs having more bulk than soil legs, so as not to fall through the environment and require additional torque resulting from dragging a partially 'submerged' leg through the gravel.
-Expect streamlined, thin design.
-legs fall to the bottom of the environment container
-large torque penalties regardless of morphology
-local minima.
Difficulty of environments:
-Gravel converges in 20 generations, 8.7 to 8.1
-Fluid converges in 40 generations, 21.5 to 19.1
-Soil converges in 80 generations,
-Soil mean of 425
-Gravel mean of 2071
-Achieved a morphology similar to something human designed
-multiple contact points in the environment
-Redundant material
-parts of splines that add useful structure or increase surface contact.
-suggests an alternate representation may be preferred.