Worked on the huggingface/lerobot repository to enhance the stability of diffusion policy training in distributed processing contexts. Addressed a critical edge case by fixing input batch formatting when handling a single observation step, which previously caused training-time errors if n_obs_steps was set to one. This solution ensured that training could proceed reliably in DP-enabled runs and established a foundation for future multi-step observation support. The work involved close collaboration with other contributors and was implemented using Python, leveraging deep learning and machine learning techniques with PyTorch. The focus was on robust engineering to improve reliability rather than adding new features.