Virtual-to-Real Motion Adaptation Using Genetic Algorithm for Multi-DOF Disaster Response Robots

For disaster response robots to work safely and effectively, it is crucial to assess task executability before execution, as failures or accidents, such as rollovers, at disaster sites could cause secondary disasters. Thus, motion learning through trial and error in such environments is unacceptable. Moreover, disaster sites are highly diverse and unpredictable, making traditional pre-installed control methods ineffective due to surface slippage and unstable terrain. In this study, we propose an environment-adaptability enhancement system that enables efficient motion learning in a virtual space for multi-degree-of-freedom disaster response robots. The process begins with the robot using a depth sensor to recognize the real environment and then reconstructs it in a virtual space. The robot subsequently attempts to complete the assigned task and calculates its confidence level in successful execution. If the task fails, the robot uses a genetic algorithm to efficiently optimize control parameters using multiple virtual robots. Finally, the learned control method is applied to the real environment. Experiments verified the robot’s ability to adaptively climb higher, uneven, and slippery steps using parameters designed for standard step climbing. The results showed that the proposed system can assess task executability via confidence levels, learn viable motions in a virtual space, and successfully perform adaptive step climbing in the real-world settings.