Worked on the facebookresearch/momentum repository to enhance the numerical stability of quaternion-to-Euler angle conversions, a core operation in gradient-based training pipelines. Focused on refactoring the quaternion_to_xyz_euler function in PyTorch, the work introduced a helper method to manage near-zero denominators, thereby improving the robustness of gradient computations. The implementation aligned with established stability patterns from real_lbs_pytorch, ensuring consistency across related codebases. Leveraging expertise in Python, PyTorch, and quaternion mathematics, the changes addressed a critical bug affecting downstream model reliability. This targeted fix contributed to more stable optimization processes without introducing new features, reflecting a depth of focus on numerical precision.