Embodied Evolution of Locomotion in Modular Robots
Control parameters tuned in simulation often don't transfer well to real robots due to the reality gap. To address this, we employ embodied evolution, evaluating candidate controllers directly on the physical robot.
Highlights Video
Overview
The controller uses a simple sinusoidal oscillator for each module. A Genetic Algorithm (GA) optimizes the amplitude, offset, and phase-shift for each module's controller, along with a common frequency parameter.
Each candidate solution (set of oscillator parameters) is tested on the physical robot. The fitness is measured as the absolute distance traveled by the robot in a fixed time, tracked using an overhead webcam and a color marker on the robot.
Methodology
- Sinusoidal oscillator controller per module
- Genetic Algorithm for parameter optimization
- Physical robot evaluation (no simulation)
- Visual tracking for fitness measurement
Why Embodied Evolution?
The "reality gap" between simulation and physical robots often causes controllers that work well in simulation to fail on real hardware. By evolving controllers directly on physical robots, we bypass this problem entirely.