Domain adaptation and long-term maintenance of invisible maps to enable low-cost autonomous navigation of robots in urban environments

点击翻译按钮获取中文摘要

This research proposes the use of "invisible maps" to enable reliable and low-cost localization on existing robots in dense crowds and dynamic environments. A major challenge to making this technology widely adopted is that the memory/computational cost required for large-scale mapping can be prohibitive. An additional challenge when considering the long-term deployment of robots is map maintenance.This year, extensive data collection has been performed using the devices developed the prior year. This data will be compared to similar data to be acquired the next year, to enable long-term performance comparisons, in particular localization accuracy degradation when using older maps.This year's main focus has been on the creation of large-scale and long-term maps. Using both adaptive grid mapping and data compression, a method has been developed which allows for large-scale invisible map creation even using low-spec computers. Long-term mapping has been achieved by developing an algorithm that can update the likelihood of changes per grid on a map. If the likelihood surpasses a set threshold, the system starts favoring newer information. Both methods are compatible and combined, enable large-scale, long-term mapping.