DiffuserLite: Towards Real-time Diffusion Planning

TL;DR

DiffuserLite is introduced, a super fast and lightweight diffusion planning framework, which employs a planning refinement process (PRP) to generate coarse-to-fine-grained trajectories, significantly reducing the modeling of redundant information and leading to notable increases in decision-making frequency.

Abstract

Diffusion planning has been recognized as an effective decision-making paradigm in various domains. The capability of generating high-quality long-horizon trajectories makes it a promising research direction. However, existing diffusion planning methods suffer from low decision-making frequencies due to the expensive iterative sampling cost. To alleviate this, we introduce DiffuserLite, a super fast and lightweight diffusion planning framework, which employs a planning refinement process (PRP) to generate coarse-to-fine-grained trajectories, significantly reducing the modeling of redundant information and leading to notable increases in decision-making frequency. Our experimental results demonstrate that DiffuserLite achieves a decision-making frequency of 122.2Hz (112.7x faster than predominant frameworks) and reaches state-of-the-art performance on D4RL, Robomimic, and FinRL benchmarks. In addition, DiffuserLite can also serve as a flexible plugin to increase the decision-making frequency of other diffusion planning algorithms, providing a structural design reference for future works. More details and visualizations are available at https://diffuserlite.github.io/.

Figures (7)

• Figure 1: Performance overview. We present Diffuser Lite, a lightweight framework that utilizes progressive refinement planning to reduce redundant information generation and achieves real-time diffusion planning. Diffuser Lite significantly outperforms predominant frameworks, Diffuser and DD, regarding scores, inference time, and model size on three popular D4RL benchmarks. The decision-making frequency of Diffuser Lite achieves $\bm{122.2}$Hz, which is $\bm{112.7}$ times higher than predominant frameworks.
• Figure 2: Comparison of one-shot planning (top) and PRP (down) on Antmaze. The former directly generates plans with a temporal horizon of $129$. The latter consists of three coarse to fine-grained levels with temporal horizons of $0$-$128$, $0$-$32$, and $0$-$8$, and temporal jumps of $32$, $8$, and $1$, respectively. The visualization in the figure illustrates the x-y coordinates of $100$ plans. It shows that one-shot planning exhibits a significant amount of redundant information and a large search space. In contrast, PRP demonstrates better plan consistency and a smaller search space.
• Figure 3: Overview of Diffuser Lite. Observing the current state $o_t$, level $0$ of Diffuser Lite fixes $o_t$ as $o_0$ and generates multiple candidate trajectories. A critic is then used to select the optimal one, in which $o_{I_0}$ is then passed to the next level as its terminal $o_{H_1-1}$. The plan refinement process continues iteratively until the last level with a temporal jump of $I_{L-1}=1$. Finally, the action $a_t$ to be executed is extracted using an inverse dynamic model $a_t=h(o_0,o_1)$.
• Figure 4: Runtime and performance comparison in FrankaKitchen. The y-axis represents the number of completed tasks (maximum of $4$), and the x-axis represents the required wall-clock time. Task success rates are presented in colored circles. All results are averaged over 250 rollouts. Diffuser Lite demonstrates significant advantages in both wallclock time and success rate.
• Figure 5: Part of selected benchmarks. From left to right, they are HalfCheetah, Hopper, Walker2d, FrankaKitchen, Robomimic, and Antmaze.