Most reinforcement learning-based humanoid motion studies emphasize full-body imitation, limiting selective correction of abnormal joints. This study proposes a reward function design that corrects abnormal joint behavior while preserving motion style, using two approaches: periodic positional targets and pre-trained joint angle references. Applied separately, both guided the agent to recover natural swing motion. Experiments in a physics-based simulation showed improved joint mobility and corrected gait patterns. The results highlight that targeted correction is achievable with imperfect motion data through reward design alone, with potential applications in rehabilitation simulations and user-feedback systems in virtual environments.

Text-to-motion generative models suffer from the long-term dependency problem, where it becomes difficult to maintain the context of text instructions as the motion length increases. Also, current MoCap datasets include only predefined actions and fail to reflect diverse individual styles. To address these limitations, we introduced MoPriC, a two-stage motion composition framework that produces sequential motions from elementary motion primitives guided by text descriptions. We also presented DancePrimitives, a new dataset of collected motion primitives to capture the semantics of each unit motion.

We present a real-time dance control system that enables users to manipulate complex full-body movements through simple hand gestures. Our system demonstrates movement modification capabilities using consumer-grade hardware, requiring only a webcam and IMU sensor for gesture capture. This research demonstrates the potential of accessible hardware for intuitive avatar control, providing immediate and responsive interaction in digital environments.

We conducted a study utilizing a lightweight generative network to create flying motions. The existing datasets used for training did not include any data on flying motions. Therefore, we selected certain classes from the existing motion datasets and transformed these motions to resemble flying actions. By training the existing generative network with the modified dataset, we were able to generate motions that closely resemble flying. The results of this study demonstrate the potential for generating flying motions. The generation of flying motions for human avatars is expected to be a critical technology not only in 3D animation or game industry but also in virtual environments, enabling users to experience various activities through their avatars.

Generating motions based on audio using deep learning has been studied steadily. However, previous research has mainly focused on speech-driven 3D gesture generation and music-driven 3D dance motion generation. We aim to generate 3D motions for specific scenarios, such as conducting. To address the challenge of lacking existing training datasets, we constructed a multi-modal 3D conducting motion dataset, which containing 1.43 hours and is, to our knowledge, a small-scale dataset. Furthermore, we propose a novel approach that uses transfer learning with a model pre-trained on a speech gesture dataset to generate 3D conducting motions. We evaluate the generated motions both with and without transfer learning, using quantitative and qualitative metrics. Our results show that the proposed method improves performance in both aspects compared to the baseline without transfer learning.