Adaptive Robot Traversability Estimation Based on Self-Supervised Online Continual Learning in Unstructured Environments

Traversability estimation is a core function for robot navigation in off-road unstructured environments and diverse research results have been published so far. One of the recent approaches is using the self-supervised learning (SSL) technique. SSL has been focused on as a breakthrough technique for situations where environments keep changing and thus traversability estimation is a challenging task. However, most of the research efforts based on SSL have several limitations: (i) they operate in an offline manner that is vulnerable to the domain distribution shift and therefore, they cannot be adaptive to the current navigation environment; and (ii) they do not take into consideration the aleatoric uncertainty of the dataset which is particularly critical in unstructured environments. In this letter, we propose an adaptive robot traversability estimation framework that considers the current navigation environment based on self-supervised online continual learning. In addition, we propose an algorithm called experience replay with uncertainty, which considers the aleatoric uncertainty of the dataset while training the traversability estimation model, thus enabling our framework to robustly estimate robot traversability. We validate our methods in various real-world environments using the Clearpath Husky robot and evaluate that our methods show better navigation performance than offline learning and rule-based methods. Moreover, we also evaluate that the proposed algorithm based on experience replay with uncertainty performs better for the benchmark dataset (ImageNet, CORe50) than the baseline algorithms.