Safe perception for autonomous navigation in natural environments at the service of agroecology – AgriVia

The progress of robotics has profoundly facilitated many of the tedious and repetitive tasks of our daily lives. This is true in activities where robots evolve in structured environments, while significant needs emerge in open and natural environments. This is the case in agriculture, a sector where current environmental and societal constraints require us to rethink production tools. Indeed, in an agroecological approach, the necessary reduction of phytosanitary products, as well as the use of bio-control products, leads to an increase in the precision of the tasks to be accomplished, and the frequency of interventions. However, the lack of manpower and the arduousness of the tasks to be carried out in high production are constraints that lead to an unsustainable burden for farmers.
Agricultural robotics, therefore, stands out as a lever of agroecology and benefits from an important national dynamic in which this proposal is included. Thus, LabCom AgriVia is taking full advantage of these issues by proposing new approaches aimed at bringing safety and precision to agricultural robots through an enriched perception of the environment and its interpretation. Relying on INNODURA TB's expertise in vision and data science, and on the research work of INRAE's Romea team in agricultural robotics, AgriVia will conduct research for the development of a safe and certified perception system for the autonomous navigation of agricultural robots. Indeed, a major obstacle in agricultural robotics is linked to the fact that most safety sensors only measure distances, and are therefore only interested in the geometry of the environment. This leads to many untimely machine stops, which then require human action to be restarted. The aim of this LabCom is to use different technologies, in particular Artificial Intelligence (AI), to provide a safe and flexible detection of the robot's environment by differentiating between traversable and nontraversable zones, depending on the type of robot and the task to be performed. To do so, the project will aim at identifying the different elements of the scene (field, road, human, etc.), their deformability or the type of vegetation, in order to adapt the robot's behavior to the evolution context and to secure the movement of people in this agricultural environment.
AgriVia's work is planned to be progressive and integrative. Indeed, these developments are focused on experimental validation and hardware and software integration throughout the LabCom. It is planned to address the problems of analysis and recognition of scenes based on several means of perception, capitalizing on existing work and expertise of partners in agricultural robotics, AI, and perception methods. This will allow within 2 years the realization of a first multi-sensor software and hardware model dedicated to the recognition of the environment and will provide learning databases allowing to develop within 4 years approaches to adapt the behavior of agricultural robots according to the context of evolution and the type of environment. The challenge is to act in a discriminating way on vegetation and soil while guaranteeing a high level of safety for people and high precision in the agricultural task performed. In order to comply with these requirements, the approaches developed must be capable of self-diagnosis in order to detect any anomalies. The objective is to qualify the hardware components, analyze the failure modes, and develop tools to verify the proper functioning of the system. Finally, AgriVia's ambition is to move towards the certification of the developed system for the safe and flexible detection of the robot's environment.