Robotic manipulation policies are increasingly empowered by large language models (LLMs) and vision-language models (VLMs), leveraging their understanding and perception capabilities. Recently, the security of robotic manipulation tasks has been extensively studied, with backdoor attacks drawing considerable attention due to their stealth and potential harm. However, existing backdoor efforts are limited to simulators and struggle to poisoning third-party commercial VLM-based implementations in real-world robotic manipulation. To address this, we propose TrojanRobot, embedding a backdoor module into the modular robotic policy via backdoor relationships to manipulate the LLM-to-VLM pathway and compromise the system, with our vanilla design employing a backdoor-finetuned VLM to serve as the module. To enhance attack performance, we further propose a prime scheme by introducing the concept of LVLM-as-a-backdoor, which leverages in-context instruction learning (ICIL) to control large vision-language model (LVLM) behavior via backdoor system prompts. Moreover, we develop three types of prime attacks—permutation, stagnation, and intentiona-—achieving flexible backdoor attack effects. Extensive physical-world and simulator experiments on 18 real-world manipulation tasks and 4 VLMs verify the superiority of proposed TrojanRobot.