Evolving Sensor Layouts Under Survival and Locomotor Constraints

We introduce a compact genotype for evolving directional sensor layouts in embodied artificial agents. The representation separates the number of sensors from their angular allocation: a scalar sensor count specifies the number of rays, while an inverse cumulative-distribution mapping converts quantiles into ordered ray headings across an evolvable field of view. This construction permits sensor-count mutation, continuous changes in angular span, and smooth foveal or peripheral concentration without explicitly mutating an unordered list of sensor coordinates. We evaluate the representation in a two-dimensional obstacle arena containing a moving target. Agents control steering and throttle through a neural network receiving only normalized obstacle-distance measurements from their evolved rays. An agent is successful if it reaches the target before colliding with an obstacle or leaving the arena; no distance, speed, path-length, survival-time, or target-proximity reward is used. We compare four conditions over seven independent 50,000-generation trials: mutable sensors without a ray-dependent locomotor cost, mutable sensors with a 3.5% speed penalty per additional ray, and fixed arrays of four or eight rays whose layouts remain evolvable. Free sensing produced substantially larger and more variable terminal sensor arrays than the cost-constrained condition (17.52 ± 7.16 versus 8.03 ± 3.03 rays). The cost condition also produced more stable terminal success (96.0% ± 4.4% versus 85.7% ± 29.5%). Fixed fourand eight-ray systems achieved nearly identical success rates (98.9% and 99.1%, respectively), despite adapting through different visual spans and angular densities. Across replicate trials, the sensor-cost condition did not converge to a single universal morphology: viable endpoints ranged from approximately three to twelve rays. These results support the inverse-CDF genotype as a useful experimental representation for studying sensor morphology, while emphasizing that the observed morphologies are task- and implementation-dependent evolutionary outcomes rather than biologically optimal designs.