Search papers, labs, and topics across Lattice.
The paper introduces an iterative closed-loop framework for generating diverse and challenging motion data to improve humanoid control policies. This framework automates motion data generation with rich action semantics and iteratively increases difficulty based on physical metrics and objective evaluations. Experiments on the PHC single-primitive tracker demonstrate a 45% reduction in failure rate on a test set using only 1/10th of the AMASS dataset size, indicating improved policy performance and data efficiency.
Generate superhuman humanoid motion data at scale by closing the loop between policy performance and data difficulty, surpassing limitations of fixed datasets.
Physics-based humanoid control relies on training with motion datasets that have diverse data distributions. However, the fixed difficulty distribution of datasets limits the performance ceiling of the trained control policies. Additionally, the method of acquiring high-quality data through professional motion capture systems is constrained by costs, making it difficult to achieve large-scale scalability. To address these issues, we propose a closed-loop automated motion data generation and iterative framework. It can generate high-quality motion data with rich action semantics, including martial arts, dance, combat, sports, gymnastics, and more. Furthermore, our framework enables difficulty iteration of policies and data through physical metrics and objective evaluations, allowing the trained tracker to break through its original difficulty limits. On the PHC single-primitive tracker, using only approximately 1/10 of the AMASS dataset size, the average failure rate on the test set (2201 clips) is reduced by 45\% compared to the baseline. Finally, we conduct comprehensive ablation and comparative experiments to highlight the rationality and advantages of our framework.
