Reconfigurable intelligent surfaces (RISs) has been proven to be a promising technology to improve the performance of symbiotic radio (SR) systems. To take full advantage of RIS-assisted communications, it is crucial to obtain accurate channel state information for high-gain beamforming. However, the statistical distribution of the cascaded channels in a RIS-assisted SR system is complex, which gives rise to a huge challenge for channel estimation tasks. To address this issue, we propose a joint optimization method of channel estimation for the RIS-assisted SR system with pilot design based on deep learning (JPCGAN) in this paper. The proposed scheme uses a fully connected network for pilot design and develops a conditional generative adversarial network (cGAN) to predict the true channels. cGAN is especially helpful in trainning the network model properly and enhance the system’s robustness with an adaptive loss function. It can be obtained by simulation that compared with current methods, the JPCGAN scheme proposed in this paper can significantly improve the accuracy of channel estimation under low signal-to-noise radios.