With interactive video streaming becoming increasingly popular, it also poses stringent requirements on the tail latency of network transport. However, achieving consistently low latency is challenging under the highly-variable wireless scenario. Existing transport mechanisms rely on network feedback (e.g., ACK) and can only react to network degradation after a time delay of RTT. When the network latency inflates abruptly, the first packets that experience the network degradation and the subsequent packets within that RTT suffer, resulting in a large portion of tail latency. To save these packets, this paper proposes Orchid that actively accelerates the in-flight packets which face deadline missing risk, thus ensuring a robust latency performance against RTT inflations in the network. In specific, Orchid monitors the in-flight video frames in fine granularity and duly duplicates video frames on multipath in a lightweight manner, to guarantee a consistently low latency against fluctuating wireless networks. Our evaluation shows that Orchid could reduce tail latency by 40× with less than 4% redundant bandwidth overhead compared to the vanilla multipath in real-world traces. The real-world deployment of Orchid in a large-scale cloud gaming service also demonstrates up to 17× fewer video stalls with a 2% bandwidth cost.