In ObjectGoal navigation (ObjectNav), agents must locate specific objects within unseen environments, requiring effective observation, prediction, and navigation capabilities. This study found that traditional methods looking only for prediction accuracy often compromise on computational efficiency. To address this, we introduce "Skip-SCAR," a modular framework that enhances efficiency by leveraging sparsity and adaptive skips. The SparseConv-Augmented ResNet (SCAR) at the core of our approach uses sparse and dense feature processing in parallel, optimizing both the computation and memory footprint. Our adaptive skip technique further reduces computational demands by selectively bypassing unnecessary semantic segmentation steps based on environmental constancy. Tested on the HM3D ObjectNav datasets, Skip-SCAR not only minimizes resource use but also sets new performance benchmarks, demonstrating a robust method for improving efficiency and accuracy in robotic navigation tasks.

翻译：在目标导向导航任务中，智能体需要在未见过的环境中定位特定物体，这要求其具备有效的观测、预测与导航能力。本研究发现，传统方法仅追求预测精度，往往在计算效率上做出妥协。为此，我们提出“Skip-SCAR”——一种利用稀疏性与自适应跳过来提升效率的模块化框架。该方法的核心是稀疏卷积增强残差网络，它并行处理稀疏与稠密特征，从而优化计算量与内存占用。我们的自适应跳过技术进一步降低了计算需求，其依据环境稳定性有选择地跳过不必要的语义分割步骤。在HM3D ObjectNav数据集上的测试表明，Skip-SCAR不仅减少了资源消耗，同时创造了新的性能标杆，为提升机器人导航任务的效率与精度提供了一种鲁棒的方法。