Artificial Intelligence (AI) agents are required to learn from their surroundings and reason about what has been learned to make decisions, act in the world, or react to various stimuli. The latest Machine Learning (ML) has adopted mostly a pure sub-symbolic learning approach. Using distributed representations of entities, the latest ML performs quick decision-making without building a comprehensible model of the world. While achieving impressive results in computer vision, natural language, game playing, and multimodal learning, such approaches are known to be data inefficient and to struggle at out-of-distribution generalization. Although the use of appropriate inductive biases can alleviate such shortcomings, in general, sub-symbolic models lack comprehensibility. By contrast, symbolic AI is based on rich, high-level representations of the world that use human-readable symbols. By rich knowledge, we refer to logical representations which are more expressive than propositional logic or propositional probabilistic approaches, and which can express knowledge using full first-order logic, including universal and existential quantification ,arbitrary -ary relations over variables,e.g. ,and function symbols,e.g. father ,etc. Symbolic AI has achieved success at theorem proving,logical inference,and verification. However, it also has shortcomings when dealing with incomplete knowledge. It can be inefficient with large amounts of inaccurate data and lack robustness to outliers. Purely symbolic decision algorithms usually have high computational complexity making them impractical for the real world. It is now clear that the predominant approach to ML, where learning is based on recognizing the latent structures hidden in the data, is insufficient and may benefit from symbolic AI [17]. In this context, neurosymbolic AI, which stems from neural networks and symbolic AI, attempts to combine the strength of both paradigms (see [16, 40, 54] for recent surveys). That is to say, combine reasoning with complex representations of knowledge (knowledge-bases, semantic networks, ontologies, trees, and graphs) with learning from complex data (images, time series, sensorimotor data, natural language). Consequently, a main challenge for neurosymbolic AI is the grounding of symbols, including constants, functional and relational symbols, into real data, which is akin to the longstanding symbol grounding problem [30].