Social robots constitute a significant component of therapeutic mediation tools used with children diagnosed with autism spectrum disorder (ASD). Predominantly anthropomorphic, these robots are designed to facilitate interaction by relying on the simplicity of their expressive modalities, achieved through simplified and predictable behaviors. Their interactional simplicity makes them particularly valuable tools for healthcare professionals. Numerous studies highlight their potential benefits in supporting joint attention, imitation, and emotion recognition. Are anthropomorphic robots, however, truly an adequate means of addressing children’s difficulties? In this article, we show that although the anthropomorphic approach currently dominates the field, it presents several limitations that must be addressed. These limitations concern the cost and accessibility of such devices, their acceptability among children and professionals, their limited adaptability to the situated practices of care, as well as the additional technological workload they impose on caregivers. Drawing on an examination of social robotics applied to autism and its limitations, this article proposes to broaden the scope of inquiry by considering a conception of the social that is not restricted to face-to-face interaction but instead encompasses its multiple dimensions. As with any technical object, the design of mediation robots is shaped by networks of human, institutional, and symbolic relations that determine their uses and their effects. Based on a review of the main existing devices and on fieldwork conducted with healthcare professionals, we identify two principal models of mediation. We argue that these models do not always align with care practices and with the situated forms of knowledge on which such practices rely. We propose to explore a complementary pathway in mediation robotics. This approach is being developed within the framework of the Médiations Robotiques en Soins de Santé (MR2S) project. It is grounded in the design of non-anthropomorphic, softer, simpler, and cheaper technical devices that build upon the experience of healthcare professionals. Our approach pays particular attention to an aspect that remains relatively unexplored in social robotics applied to autism: sensoriality. The purpose of the project is to design robotic objects capable of enriching the therapeutic relationship while being sustainably integrated into existing care practices.