Under the term ultra-reliable low-latency communications (URLLC), the 5th generation (5G) of cellular networks promises to deliver 99.999 % of sent data packets within one millisecond. Such ambitious figures are demanded by industrial use cases with closed-loop motion control, for example. But as of today, no commercially available wireless system is known to actually meet these requirements. There are three major limits to reliable wireless communications: (1) sudden loss of signal power between transmitter and receiver (fading), (2) third-party interference from other wireless devices, and (3) propagation-related signal distortion independent of noise and third-party interference. This thesis focuses on problems (1) and (3). It presents practical insights and latency-friendly solutions to improve reliability using frequency diversity. First, multi-connectivity with diversity combining on the physical layer is evaluated for IEEE 802.11 wireless local area networks by means of Monte-Carlo simulations assuming various fading models. With increasing number of uncorrelated links, multi-connectivity achieves much lower error rates than a single link. Joint decoding based on distributed turbo coding is found to outperform the established combining schemes selection combining and maximum ratio combining when considering receiver imperfections in the presence of doubly-selective fading. This was not expected from theoretical work and shows the importance of studies going beyond simplified analytical models. To better understand the wireless propagation conditions in practice, high-resolution channel measurements are captured at a Bosch factory hall and analyzed with a focus on reliability. Metallic surfaces in the environment are found to lead to a small path loss over distance, but also to severe fading. Fortunately, the coherence bandwidth is small which promotes the use of frequency diversity. To overcome the low spectral efficiency of multi-connectivity, schemes for dynamically allocating bandwidth to many users in real-time are developed and evaluated in Monte-Carlo simulations using the factory hall measurements. A low-complexity algorithm for channel-aware allocation is proposed and found to perform very close to a best-case bound. It is shown that computational complexity and sounding overhead can be reduced with negligible loss of reliability by increasing subchannel width and channel state update interval in accordance with coherence bandwidth and time. In conclusion, the insights on multi-connectivity and the channel-aware allocation algorithm are believed to be valuable contributions to fulfill the URLLC promise.