Training and Evaluating Diffusion Policies with Long Context Lengths

This work presents a detailed evaluation of imitation learning policies with varying context lengths, demonstrating that longer context can improve performance on tasks requiring memory without the brittleness previously suggested. Using a UNet architecture with cross-attention for denoising, the study shows that single-task policies can achieve high success rates even with naive context length scaling. Additionally, a novel training algorithm is introduced that enables joint training across multiple context lengths, thereby reducing sample complexity for long-context learning, which is significant for practitioners aiming to enhance robotic manipulation capabilities.