Biological systems can generate robust and complex behaviors through limited local interactions in the presence of large amount of uncertainties. Inspired by biological organisms, we have proposed a gene regulatory network (GRN) based algorithm for self-organizing multiple robots into different shapes. The self-organization process is optimized using a genetic algorithm. This paper focuses on the empirical analysis of robustness of the self-organizing multi-robot system to the changes in tasks, noise in the robot system and changes in the environment. We investigate the performance variation when the system is optimized for one shape and then employed for a new shape. The influence of noise in sensors for distance detection and self-localization on the final positioning error is also examined. In case of a complete self-localization failure, we introduce a recovery algorithm based on trilateration combined with a Kalman filter. Finally, we study the system's performance when the number of robots changes and when there are moving obstacles in the field. Various simulation results demonstrate that the proposed algorithm is efficient in shape formation and that the self-organizing system is robust to sensory noise, partial system failures and environmental changes.