Evolving Gate and Residual Reflexes for Dynamic Obstacle Avoidance in a Hybrid Robot-Arm Controller
Computer Science Journal of Moldova · 2026-11 · 34 (2) · 202-215
DOI: 10.56415/csjm.v34.09 ↗Kurz gefasst
Fachartikel in der Zeitschrift Computer Science Journal of Moldova, 2026-11. Verfasst von Andrei Darii und 2 weiteren Autor:innen. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-SA 4.0).
Abstract (Originalsprache)
We examine a hybrid controller for a simulated WLkata Mirobot. In this controller, a damped least-squares inverse-kinematics planner reaches targets, while a neuroevolved neural reflex supplies only obstacle-avoidance corrections. We compare two reflex architectures: a scalar braking gate that can only slow the planner, and a vector residual policy that can sidestep. Each architecture is evolved either from scratch (cold) or from a supervised demonstrator (warm). Across 256 held-out seeds evaluated on both the training speed band (0.3-0.5 m/s) and a harder out-of-distribution band (0.0-1.5 m/s), the two architectures reach comparable target counts. The residual policy does so with roughly one-half to one-third of the collision rate of the brake-only gate, indicating that it uses lateral avoidance rather than braking alone. Warm-starting does not improve the peak target count over cold-start evolution. Its main benefit is reliability, as it removes the single catastrophic cold-start failure and sharply reduces between-run variance. Held-out rankings also diverge from training fitness, underscoring the need to evaluate on unseen seeds rather than on the evolutionary objective. These findings are based on a small number of evolved runs per cell on a single simulated platform and one obstacle scenario. They should therefore be read as indicative rather than statistically conclusive.
Zitieren
APA
Darii, A., Nistor, M. S., & Pickl, S. (2026). Evolving Gate and Residual Reflexes for Dynamic Obstacle Avoidance in a Hybrid Robot-Arm Controller. Computer Science Journal of Moldova, 34(2), 202-215. https://doi.org/10.56415/csjm.v34.09
Vancouver
Darii A, Nistor MS, Pickl S. Evolving Gate and Residual Reflexes for Dynamic Obstacle Avoidance in a Hybrid Robot-Arm Controller. Computer Science Journal of Moldova. 2026;34(2):202-215. doi:10.56415/csjm.v34.09
BibTeX
@article{Darii2026,
author = {Darii, Andrei and Nistor, Marian Sorin and Pickl, Stefan},
title = {Evolving Gate and Residual Reflexes for Dynamic Obstacle Avoidance in a Hybrid Robot-Arm Controller},
journal = {Computer Science Journal of Moldova},
year = {2026},
volume = {34},
number = {2},
pages = {202--215},
doi = {10.56415/csjm.v34.09}
}
EUVIMED