Evolutionary Developmental Robotics

Guettas, Chourouk (2026) Evolutionary Developmental Robotics. Doctoral thesis, Université Mohamed Khider - BISKRA.

Text Thesis__Chourouk__Guettas__.pdf Download (4MB)

Abstract

The field of robotics has long grappled with the challenge of creating adaptive learning systems that efficiently acquire new skills while maintaining previously learned capabilities. Traditional approaches rely on hand-engineered features and rigid architectures that struggle to generalize across diverse environments. While deep reinforcement learning has shown success, these methods suffer from sample inefficiency and require extensive computational resources that limit real-world deployment. This research introduces two novel bio-inspired frameworks that translate natural regulatory mechanisms into computational advantages for robotic learning. The first contribution presents a Bio-Inspired Adaptive Rate Modulation (BIARM)framework that coordinates four computational modules corresponding to key neurotransmitter functions: dopamine for reward-based adaptation, serotonin for stability control, norepinephrine for attention gating, and acetylcholine for knowledge integration. Unlike existing methods that optimize single factors independently, BIARM implements coordinated multi-factor adaptation using organizational principles derived from biological neuromodulation. The second contribution explores Gene Regulatory Networks (GRNs) for visionbased robotic control, representing robot states as gene expression levels and using evolutionary optimization to discover regulatory relationships that map sensory inputs to motor commands. This approach leverages fundamental biological properties including bounded dynamics, sparse connectivity patterns, and natural temporal reasoning without explicit memory modules. Experimental validation demonstrates significant effectiveness. BIARM achieves performance improvements ranging from 1.06% to 155.94% over baseline Progressive Neural Networks across four OpenAI Gymnasium environments, while the GRN-based controller achieves 57.5% success rate in the KukaDiverseObjectEnv benchmark with up to 13.7× reduction in training time compared to established reinforcement learning baselines. These results demonstrate that biologicallyinspired principles can offer algorithmic advantages for autonomous system design

Actions (login required)

View Item