Lite-DRTG: A Lightweight Diffusion-Based Trajectory Generation for Real-Time Mapless Navigation

The need for mapless navigation is growing as demand for autonomous mobile robots surges in last-mile delivery, search-and-rescue, and logistics. The trajectory generation from onboard sensor data without prior maps helps robots to plan safe, long-range routes in an outdoor environment where construction and sudden obstacles rapidly invalidate prebuilt maps. However, mapless trajectory generation remains challenging in complex outdoor environments that are highly dynamic, uncertain, and unstructured. Recently, diffusion-based trajectory generation (DTG) has emerged as a state-of-the-art paradigm for mapless navigation. However, the performance and inference efficiency of DTG are highly contingent upon the architecture of the denoiser backbone. Although richer representations can enhance route traversability and safety, they typically incur substantial computational overhead of inference. In this paper, we propose Lite-DRTG, which employs a denoiser backbone called AdaLite-UNet, a lightweight one-dimensional (1D) UNet, designed for real-time mapless navigation. AdaLite-UNet achieves computational efficiency by fundamentally restructuring and lightweighting the UNet architecture, and by incorporating feature-wise linear modulation-based adaptive group normalization across all layers. Experimental results show that Lite-DRTG (with AdaLite-UNet) achieves significant performance gains: compared to the convolutional recurrent neural network-based DTG baseline, traversability improves by 22.1%, and our model outperforms the much heavier DTG (UNet) variant on most metrics. Notably, Lite-DRTG is 81.3% faster and 74.1% smaller than the DTG (UNet) variant while also maintaining a smaller distance error.