Neuroevolution for Actively Controlled Virtual Characters
Abstract: Physics-based character animation offer an attractive alternative to traditional animation techniques, however, physics-based approaches often struggle to incorporate active user control of these characters. This thesis suggests a different approach to the problem of actively controlled virtual characters. The proposed solution takes a neuroevolutionary approach, using HyperNEAT to evolve neural controllers for a simulated eight-legged character, a previously untested character morphology for this algorithm. Using these controllers this thesis aims to evaluate the robustness and responsiveness of a control strategy that changes between them based on simulated user input. The results show that HyperNEAT is quite capable of evolving long walking controllers for this character, but also suggests a need for further refinement when operated in tandem.
Result
It is highly recommended to read the report before watching the result as this gives significantly more context to the videos. The report can be found in the repository linked above.
The videos on this page are displayed in the same order as they appear in the report. This is true for all videos, except for the videos marked others
as these are not included in the report itself.
All of the videos are running at half the speed in order (30 fps) to empathize the motions of the character.
Standing
Standing 01
Standing 01 - alternative
Standing 02
Alternate champion
Champion
Final Network of the Standing Champion
This video displays the final network (phenotype) produced by CPPN using the substrate. See the report for details.
Walking
Walking 01
Walking 02
Walking 03
Walking 04
Walking 05
Walking 06
Walking 07
Others
These next videos did not have images series of them included in the report.
Champion
Walking Champion
Champion throughout the generations
This video displays how the champion changed from its first generation until its last. This video was not included in the report.
Final Network of the Walking Champion
This video displays the final network (phenotype) produced by CPPN using the substrate. See the report for details.