DiffuserLite: Towards Real-time Diffusion Planning

Zibin Dong, Jianye Hao, Yifu Yuan, Fei Ni, Yitian Wang, Pengyi Li, Yan Zheng
Computer Science, Artificial Intelligence, Artificial Intelligence (cs.AI)
2024-01-27 00:00:00
Diffusion planning has been recognized as an effective decision-making paradigm in various domains. The high-quality conditional generation capability of long-horizon trajectories makes it a promising research direction. However, existing diffusion planning methods suffer from low decision-making frequencies because of the expensive iterative sampling cost. To address this issue, we introduce DiffuserLite, a fast and lightweight diffusion planning framework. DiffuserLite employs a planning refinement process (PRP) to generate coarse-to-fine-grained trajectories, which significantly reduces the modeling of redundant information and leads to notable increases in decision-making frequency. Our experimental results demonstrate that DiffuserLite incurs only $0.88\%$ of the runtime cost compared to previous frameworks, achieves an average decision-making frequency of $122$Hz, and reaches state-of-the-art performance on D4RL benchmarks. In addition, our clean DiffuserLite framework can serve as a flexible plugin to enhance decision frequency in other diffusion planning algorithms, providing a structural design reference for future works. More details and visualizations are available at [project website](
PDF: DiffuserLite: Towards Real-time Diffusion Planning.pdf
Empowered by ChatGPT