GECCO 18 presentation

Real-World Evolution Adapts Robot Morphology and Control to Hardware Limitations

Tønnes F. Nygaard, Charles P. Martin, Eivind Samuelsen, Jim Torresen, Kyrre Glette

Our long term goal

Be able to do open ended evolution in dynamic environments in the real world, to adapt a robots control and morphology to unexpected events.

Background

Use of evolutionary robotics for legged robots:

Most only evolve control, but there has been some work evolving morphology for hardware robots
Some evolve morphology exclusively in software, and transfer them to the real world, as we see to the left
Others evolve exclusively in hardware, as we see in the middle, where small legged robots are assembled by a robot arm
Others robots, like we see to the right, have the ability to change its own morphology, but it is only changed during the evolution to improve speed or performance, and at the end of evolution, and we're only left with the single hand-designed morphology.
Very few examples of morphology evolution on advanced legged robots done in hardware

Vuk Vujovic, Andre Rosendo, Luzius Brodbeck, and Fumiya Iida. 2017. Evolutionary developmental robotics: Improving morphology and control of physical robots. Artificial Life (2017).
Josh Bongard. 2011, Proceedings of the National Academy of Sciences, Morphological change in machines accelerates the evolution of robust behavior

Challenges

Goals for our paper

Evolve control and morphology in
hardware on our new robot platform
using a high-level controller
Show that the system can adapt to
different hardware conditions

DyRET

Dynamic Robot for Embodied Testing

Self-reconfiguring legs

Control system

Spline-based gait

Evolutionary setup

Fitness:
- Speed (from mocap)
- Stability (from IMU)
Algorithm: NSGA-II

A full evolutionary run

Experiment

Evolutionary runs at two different
servo voltages
Investigate effects of lowering voltage

Results - fitness comparison

Results - re-evaluation

Results - individuals

Significant differences for control and morphology

Results - leg lengths

Optimal voltage

Reduced voltage

Conclusion

Reducing servo voltage severely
reduces performance
Evolutionary algorithms are able to adapt to
this and get comparable performance
at low and medium speeds
Both control and morphology is changed
to achieve this

Future work

DyRET needs your help to cross the reality gap!

https://github.com/dyret-robot/dyret_documentation