Standing the Test of Time: Retrospective and Future of World Representations for Lifelong Robotics

Accurate, informative, and scalable world representations are an essential component of highly autonomous mobile robots and have been an important topic of research for several decades. As robots become more capable, deploying in larger and more dynamic, varied environments, requires for such representations to grow apace. Handling multiple data modalities, abstraction levels, and types of information (metric, topological, semantic, objects, etc.) remains challenging — even more so in so-called lifelong settings where robots must maintain world models over extended periods of time. Over the last forty years, roboticists have used techniques from machine learning and statistics paradigms for mapping. However, none have been nearly as transformative as deep learning, and we believe we are now at an inflection point in the pace of adoption and proliferation of deep learning techniques for representing models of the world suited to robotics. Such a moment offers an opportunity for retrospection: to consider lessons from previous eras of research that have stood the test of time, to carry such lessons forward into an age of research dominated by models relying on latent representations, and to understand in hindsight the limits and blind spots of previous paradigms. Looking forward, we also hope to: make progress understanding the tradeoffs presented by newer learning and representation techniques, share and discuss new examples of state-of-the-art technical approaches for robotic mapping and modeling, and develop a shared view of the new frontier of challenges facing such systems as they are deployed in ever more challenging domains.