Based upon user-configurable modular robotics and design principles for modular playware, we developed modular robotic tiles to be used as playful, interactive tools for children with autism. The modular playware can make automatic documentation of the construction play activities by the autistic children. Using artificial neural networks for automatic classification of the individual construction practices, we may compare this classification with the diagnosis of the children, and possible obtain a supplementary diagnosis tool which is based on the autistic children's free play with the modular robotic tiles. Preliminary experiments with 7 autistic children show that the automatic neural network classification with post-processing can be done with a 100% accuracy for this small sample set, and thereby give some preliminary indications of the potential of the approach.