Creating a Framework for Application of Transferability Approach – Background – Background
Creating a Framework for Application of Transferability Approach – Background – Background
0 0
master-slides
On Github RobertKolner / master-slides
Creating a Framework for Application of Transferability Approach
Overview
How is this presentation structured?
Goals Background and Implementation Testing Results ConclusionGoals
Primary:
Develop a framework for simulation and real-life experiments on Aracna Extend the framework to be able to conduct experiments with Transferability Approach.Secondary:
Evaluate viability of Aracna as a platform for the process.Background
- Theory - EA, MOEA, Reality Gap, Transferability
- Software - Simulator based on PhysX
- Hardware - Aracna
Evolutionary Algorithms
A family of flexible optimization algorithms, typically used (in robotics) to generate controllers optimized for different goals, morphology (ER) and pathfinding.
There are many types of evolutionary algorithms:
GA - Genetic Algorithms GP - Genetic Programming EP - Evolutionary Programming ES - Evolution StrategyTypical stages of an evolutionary algorithm:
Multi-Objective Evolutionary Algorithms
- Pareto-front and domination of individuals
- NSGA-II - an effective algorithm for sorting of populations with regard to multiple objectives
An example of a multi-objective optimization run resulting in a Pareto-front:
Reality Gap
Why don't our gaits do anything on a real robot?
- Imperfections of the model in the simulator
- Simplified physics
- Unreliable hardware
- Varying environment
- ...
Transferability Approach
- Behaviours (Examples: mean height, distance, changes of orientation)
- Disparity Value
- Surrogate model
- SVM --> SVR
- Initial transfer set diversity
- Distance --> || b1 - b2 ||
- Choosing the behaviours and normalization
Background
- Theory - EA, MOEA, Reality Gap, Transferability
- Software - Simulator based on NVidia PhysX
- Hardware - Aracna
Simulator
- Developed by the ROBIN-group at UiO
- Based on NVidia PhysX
Background
- Theory - EA, MOEA, Reality Gap, Transferability
- Software - Simulator based on PhysX
- Hardware - Aracna
Aracna
Image source: http://creativemachines.cornell.edu/aracnaFeatures:
- 3D-printed structural elements complemented by AX18 servos
- Servos are located centrally in the body
- ...which leads to unconventional kinematics
Side view of Aracna's leg:
Talk about Symmetry, Friction at the tips, Torque which is difficult to compute, Search space
Testing
Goals
Primary:
Develop a framework for simulation and real-life experiments on Aracna Extend the framework to be able to conduct experiments with Transferability Approach.Secondary:
Evaluate viability of Aracna as a platform for the process.Step by step verification
The GA framework produces viable gaits Similarity of the simulated and physical model Using controllers on the real robot Comparison results with and without Transferability Approach- Fitness (movement distance, transferability)
Results
The GA framework produces viable gaits
Similarity of the simulated and physical model
Using controllers on the real robot
Comparison of gaits after 2s and 8s
Comparison of results with and without Transferability Approach
1st iteration
2nd iteration
3rd iteration
4th iteration
Why?
- Not entirely correct implementation of Transferability Approach
- Wrong choice of behaviours
- Questionable choice of SVM-parameters
- Additionally...
Behaviours used: distance moved, sum of orienatation changes, direction moved
Aracna
Small changes in controller and behaviours lead to big changes in how the gaits perform.
Conclusion
- A simple model of Aracna was created, tested and evaluated, both in simulation and reality
- Application of Transferability Approach did not yield expected results
- ...but a lot of work remains to be done
Aracna might not have been the best choice for this project.
Future work
- Improve model of Aracna
- Transfer experiments with other robots
- Test different behaviours
- Local search
Model of Aracna: