Quad-Imaginarium
收藏资源简介:
Quad-Imaginarium是由阿里巴巴集团·高德团队创建的大规模四足机器人运动数据集,旨在解决传统动物运动捕捉数据稀缺、形态不匹配等问题。该数据集包含7,488条高质量3D参考运动轨迹,总时长18.5小时,涵盖杂技性和表演性行为,每条数据均配有语言标注和人工校正的动作标签。数据集通过创新的Uni-Mo流水线生成,结合大型语言模型提出运动描述、视频扩散模型合成机器人行为、并通过身份一致性损失确保生成质量,最终转化为可部署的跟踪策略。该数据集主要应用于四足机器人表达性运动生成、强化学习策略训练等领域,为突破传统运动数据局限提供了计算驱动的解决方案。
Quad-Imaginarium is a large-scale quadruped robot motion dataset developed by Alibaba Group’s Amap Team, aimed at addressing the issues of scarcity and morphological mismatch of traditional animal motion capture data. This dataset contains 7,488 high-quality 3D reference motion trajectories, with a total duration of 18.5 hours, covering acrobatic and performative behaviors. Each entry is equipped with linguistic annotations and manually corrected action labels. The dataset is generated through an innovative Uni-Mo pipeline, which combines large language models (LLMs) to generate motion descriptions, uses video diffusion models to synthesize quadruped robot behaviors, and leverages identity consistency loss to ensure the quality of generated outputs, ultimately being converted into deployable tracking policies. This dataset is primarily applied in fields such as expressive motion generation for quadruped robots and reinforcement learning policy training, providing a computation-driven solution to break through the limitations of traditional motion data.




