Climbing Motion Generation for Humanoid Agents through Deep Reinforcement Learning with Optimization Constraints

By enabling a humanoid to perform human-like movements, it can be utilized in various applications, such as assisting humans and motion analysis. While research on tasks like walking and transporting objects with humanoids has advanced, studies involving climbing walls remain underexplored. In climbing, while the movement of the hands gripping the wall is crucial, whole-body movement is also essential for reducing the load on the hands. This study focuses on bouldering as an example where full-body coordination is critical during climbing, and employs deep reinforcement learning to train a humanoid in a simulation environment to climb while minimizing hand load. As a result, the humanoid achieved a high success rate in the given task, demonstrating that deep reinforcement learning can successfully teach climbing methods that alleviate the load on the hands. Furthermore, the study suggests that gradually increasing the weight assigned to a subtask is an effective strategy when simultaneously learning two tasks.