We address small jellyfish detection in underwater robotics by proposing LSJD-YOLO, a lightweight object detector based on an enhanced YOLO architecture, designed for efficient edge deployment. To address the limited representation of small targets in existing datasets, we introduce two new benchmarks: R-Jellyfish and Small Jellyfish. LSJD-YOLO integrates a Small Target Detection Head (STDH) to improve detection capability on small objects and a Lightweight Multi-scale Feature Fusion (LMFF) module to reduce channel dimensions and model complexity. Filter pruning is further applied for model compression. Evaluated on a Raspberry Pi with inference acceleration, the model achieves real-time performance. On R-Jellyfish, it improves $A P_{s}$ by 4.2 %, reduces parameters by 64.7 %, and model size by 60.2 %. On Small Jellyfish, $A P_{s}$ increases by 1.6 % with 63.6 % fewer parameters and 59.2 % smaller size. With acceleration, detection speed improves by 247.4 % and 190.5 %, respectively, demonstrating strong potential for real-time underwater applications. The dataset and code are publicly available at: https://github.com/yzl12345/LSJD-YOLO