Progress in robotic intelligence: the role of foundation models

Purpose The purpose of this paper is to provide an insight into how foundation models are transforming robotic intelligence and capabilities. Design/methodology/approach Following a short introduction, this first discusses foundation model technologies. It then considers their robotic applications and provides examples of research and product developments. Finally, conclusions are drawn. Findings Foundation models are large-scale AI models trained unsupervised on vast amounts of unlabelled data. They allow one large-scale pretrained model to be customised for a wide range of downstream tasks with little additional task-specific data. The technology catalysed the modern era of GenAI and major classes include LLMs and VLMs. Foundation models aimed specifically at robotic applications have recently been developed. Trials with robots employing LLMs and VLMs have demonstrated hitherto unpresented functionalities, including interpreting and responding to natural language commands, recognising and manipulating hitherto unseen objects, generalising to new tasks not seen during training and interpreting their surroundings. The capabilities have been exploited in humanoids which have conducted complex industrial tasks and routine household actions. A growing number of companies are aiming to commercialise domestic humanoids and foundation models are seen as the key to achieving this. Foundation models represent a paradigm shift in machine intelligence and their application to robotics is poised to transform many sectors of the industry. Originality/value This shows how foundation models are transforming the capabilities of robots by imparting them with hitherto unprecedented levels of machine intelligence.