Generating Double Dyno Motion for Humanoid Agents in Simulated Bouldering Environment Through Deep Reinforcement Learning

Capturing human motion data is crucial for creating realistic character animation and analyzing sports movements. With the widespread availability of motion capture devices, obtaining realistic motion data has become more accessible. Meanwhile, it is still challenging to capture quality motion data of extensive movements due to bulky equipment or simply the high cost. This paper investigates the feasibility of generating realistic motion data by combining virtual humanoids and deep reinforcement learning. Specifically, we take bouldering as a target environment where it is challenging to deploy conventional motion capture devices. If motion generation can be achieved in bouldering, a potential application could involve observing exemplary motions through VR or AR, thereby enhancing the efficiency of training. We trained a humanoid to perform a double dyno in a simulated bouldering environment and compared this motion to a human performing the same move. By evaluating the realism of the motion data, we found that although the humanoid's movements differed from human movements, there were certain aspects where the humanoid's actions closely resem-bled those of humans. Furthermore, it is suggested that applying penalties for inappropriate movements can lead to more natural motions.