This paper proposes LightDP, a novel framework designed to accelerate the real-time deployment of diffusion policies, which struggles due to limited mobile device resources. LightDP addresses computational bottlenecks through two key strategies: network compression of the denoising module and reduction of required sampling steps. We conduct extensive computational analysis of existing diffusion policy architectures to identify the denoising network as a primary contributor to latency. To overcome the performance degradation associated with existing pruning methods, we introduce an integrated pruning and retraining pipeline that explicitly optimizes the model's post-pruning resilience. Furthermore, we combine pruning techniques with consistency distillation to effectively reduce sampling steps while maintaining action prediction accuracy. Experimental evaluations on standard datasets such as PushT, Robomimic, CALVIN, and LIBERO demonstrate that LightDP achieves competitive performance for real-time action prediction on mobile devices, representing a significant step toward practical deployment of diffusion-based policies in resource-constrained environments. Extensive real-world experiments demonstrate that the proposed LightDP achieves performance comparable to state-of-the-art diffusion policies.
