Emotion recognition is a strategy for social robots used to implement better Human-Robot Interaction and model their social behaviour. Since human emotions can be expressed in different ways (e.g., face, gesture, voice), multimodal approaches are useful to support the recognition process. However, although there exist studies dealing with multimodal emotion recognition for social robots, they still present limitations in the fusion process, dropping their performance if one or more modalities are not present or if modalities have different qualities. This is a common situation in social robotics, due to the high variety of the sensory capacities of robots; hence, more flexible multimodal models are needed. In this context, we propose an adaptive and flexible emotion recognition architecture able to work with multiple sources and modalities of information and manage different levels of data quality and missing data, to lead robots to better understand the mood of people in a given environment and accordingly adapt their behaviour. Each modality is analyzed independently to then aggregate the partial results with a previous proposed fusion method, called EmbraceNet+, which is adapted and integrated to our proposed framework. We also present an extensive review of state-of-the-art studies dealing with fusion methods for multimodal emotion recognition approaches. We evaluate the performance of our proposed architecture by performing different tests in which several modalities are combined to classify emotions using four categories (i.e., happiness, neutral, sadness, and anger). Results reveal that our approach is able to adapt to the quality and presence of modalities. Furthermore, results obtained are validated and compared with other similar proposals, obtaining competitive performance with state-of-the-art models.< Réduire