Gene Regulatory Networks for Enhanced Vision-Based Robot Control: A Bio-Inspired Approach.

Area of Science:

• Robotics
• Artificial Intelligence
• Computational Biology

Background:

• Deep reinforcement learning (DRL) faces challenges in sample inefficiency and long training times for vision-based robot control.
• Existing DRL methods struggle with robustness in real-world scenarios, especially under visual noise.

Purpose of the Study:

• To develop a novel, efficient, and robust robot control method inspired by biological Gene Regulatory Networks (GRNs).
• To leverage GRNs for mapping raw visual inputs to robot motor commands, overcoming DRL limitations.

Main Methods:

• Encoding robot states as gene expression levels within a GRN framework.
• Utilizing evolutionary optimization to learn GRN parameters for visual-to-motor command mapping.
• Evaluating the GRN controller on the KukaDiverseObjectEnv benchmark for object grasping tasks using RGB images.

Main Results:

• Achieved a 57.5% success rate in object grasping, outperforming baselines like Proximal Policy Optimization (PPO) and Deep Q-Learning.
• Reduced training time by 13.7× compared to PPO.
• Maintained 91.8% performance under noisy visual conditions, demonstrating robustness.

Conclusions:

• The GRN-based approach offers a computationally efficient and robust solution for vision-based robot control.
• Bio-inspired GRNs enable hierarchical control, computational efficiency, and temporal reasoning without explicit memory.
• This method presents a promising alternative to traditional DRL for complex robotic tasks.