Vision-based robot control remains a significant challenge due to the sample inefficiency and prolonged training times associated with traditional deep reinforcement learning methods. We propose a novel approach inspired by biological gene regulation, leveraging Gene Regulatory Networks (GRNs) for efficient and robust robot control. In our approach, robot states are encoded as gene expression levels, and evolutionary optimization is used to learn GRN parameters that map raw visual inputs to motor commands. We evaluate this method on the KukaDiverseObjectEnv benchmark, where robots must grasp diverse objects using only RGB images. Our GRN-based controller achieves a 57.5% success rate while reducing training time by 13.7× compared to Proximal Policy Optimization baselines. It also outperforms NEAT, standard reinforcement learning algorithms, and deep Q-learning in terms of both efficiency and performance. The controller maintains 91.8% performance under noisy visual conditions. This bio-inspired design naturally enables hierarchical control via expression cascades, computational efficiency through bounded dynamics, and temporal reasoning without explicit memory modules.
Guettas et al. (2026) studied this question.