CALIPAM - Calibrated Uncertainty in Perception and Mapping for Robust Robot Deployment in Agricultural Environments
CALIPAM is an ANR JCJC project starting in 2027.
In recent years, robotics and autonomous systems have advanced significantly, particularly in precision agriculture. Nonetheless, the broader development of robotics for open-field applications still requires considerable effort. While many robotic systems perform well in controlled environments, they often struggle in complex, unstructured, and dynamic outdoor environments. Making robotics systems robust to handle unpredicted events that can occur in their operating environment is a necessary step for their wider adoption. Autonomous navigation of unmanned ground vehicles in natural and unstructured environments still faces scientific barriers. On the one hand, some tasks, like localization and mapping, are typically addressed by probabilistic approaches, quite mature, that present the significant advantage of predictability. On the other hand, deep learning-based perception methods, such as object detection and segmentation, offer new possibilities, but lack predictability and do not provide an explicit representation of uncertainty. These limitations, coupled with the inability of training data to capture the full complexity of real-world environments, make deep learning less reliable for real-world missions. CALIPAM aims to enhance the reliability of robotic systems by combining deep-learning-based perception with probabilistic methods into an uncertainty aware world representation that enables reliable execution of high level missions. The focus is on two key areas: perception and mapping, specifically for agricultural robotics in open fields and orchards. Two scenarios drive the research: identifying uncertainties in the perception that could lead to false conclusions during monitoring tasks, and detecting situations where the uncertainty in the map during navigation tasks might cause mission failures, requiring human intervention.