SenseExpo: Efficient Autonomous Exploration with Prediction Information from Lightweight Neural Networks

TL;DR

SenseExpo introduces a lightweight autonomous exploration framework that predicts local occupancy maps using a 709k-parameter network combining GANs, Transformer, and Fast Fourier Convolution. The system operates on partial onboard observations, fuses local predictions across time and robots, and guides frontier-based exploration, all packaged as a plug-and-play ROS node for easy integration. It demonstrates state-of-the-art prediction quality on the KTH dataset, strong cross-domain robustness on HouseExpo, and substantial exploration-time reductions (up to ~77% on MRPD 1.0) compared with existing methods. The work highlights practical applicability for resource-constrained robots, and its modular ROS deployment enables widespread adoption in real-world navigation and mapping tasks.

Abstract

This paper proposes SenseExpo, an efficient autonomous exploration framework based on a lightweight prediction network, which addresses the limitations of traditional methods in computational overhead and environmental generalization. By integrating Generative Adversarial Networks (GANs), Transformer, and Fast Fourier Convolution (FFC), we designed a lightweight prediction model with merely 709k parameters. Our smallest model achieves better performance on the KTH dataset than U-net (24.5M) and LaMa (51M), delivering PSNR 9.026 and SSIM 0.718, particularly representing a 38.7% PSNR improvement over the 51M-parameter LaMa model. Cross-domain testing demonstrates its strong generalization capability, with an FID score of 161.55 on the HouseExpo dataset, significantly outperforming comparable methods. Regarding exploration efficiency, on the KTH dataset,SenseExpo demonstrates approximately a 67.9% time reduction in exploration time compared to MapEx. On the MRPB 1.0 dataset, SenseExpo achieves 77.1% time reduction roughly compared to MapEx. Deployed as a plug-and-play ROS node, the framework seamlessly integrates with existing navigation systems, providing an efficient solution for resource-constrained devices.

Figures (10)

• Figure 1: SenseExpo Framework. The Map Predictor uses odometry and global observation map obtained from the robot to produce local predicted map. The Predicted Map Server then concatenates the local predicted map and the global predicted map to form a predicted map layer. Subsequently, the Planner Server fuses the global predicted maps of each robot (if there are multi robots) into a total predicted map layer and outputs the path to the navigation point.
• Figure 2: Complete Architecture of the Map Prediction Network. The input, which has 3 channels (free, uncertain, obstacles), is processed through a U-Net model. Simultaneously, input patches are passed through a Transformer Encoder and then Fast Fourier Convolution, with the outputs concatenated to the feature maps in the U-Net.
• Figure 3: Illustration of the model architecture. (a) The spectral transform module consists of several Conv-BN-ReLU and FFT2d layers. The input is divided into local and global, which are processed through the spectral transform block and Conv-BN-ReLU block, then concatenated to form new local and global input. The output is subsequently passed into another spectral transform block and Conv-BN-ReLU block. (b) The Transformer Encoder block consists of a series of processing layers, including normalization, multi-head attention, and a multi-layer perceptron (MLP), each applied sequentially to the embedded input patches. The output is then passed through repeated layers to extract high-level features.
• Figure 4: The example of dataset. The True Map is encoded into three channels, free, uncertain and obstacles.
• Figure 5: Comparison of uncertain region distribution and free region boundary extraction between traditional and proposed methods. (a) Uncertain regions concentrated at obstacle edges in the traditional approach; (b) Boundary points extracted from the free region in the proposed method.