Agbots 2.0: Weeding Denser Fields with Fewer Robots

Agbots 3.0: Adaptive Weed Growth Prediction for Mechanical Weeding Agbots

“…However, for these robots to be employed at scale, multiagent planning strategies must solve the problem of coordination in field environments with limited observations. This work builds on past work presented at the Robotics: Science and Systems Conference 2020 [1], as well as earlier work [16], [20]. In [16] and [20], a realistic simulation environment, Weed World, was developed and used to demonstrate the feasibility of a coordinated strategy for robotic weeding of fields with a range of characteristics, with the a robot similar to TerraSentia as the basis for our simulation parameters.…”

“…In [16] and [20], a realistic simulation environment, Weed World, was developed and used to demonstrate the feasibility of a coordinated strategy for robotic weeding of fields with a range of characteristics, with the a robot similar to TerraSentia as the basis for our simulation parameters. In [1], we presented an improved planning scheme, which more effectively balanced the explore-exploit tradeoff while compiling information on unknown rows. However, in all previous work, the coordination strategy used the average values for the weed height and density of previously weeded rows for the value of weeding unobserved rows.…”

“…The simulation environment, called Weed World (shown in Fig. 3), was developed to allow large-scale simulations of coordinated weeding algorithms for multirobot planning in uncertain environments [1]. This environment incorporates a realistic weed growth model (described in Section II-B), as well as a framework for multiagent collaboration, which enables a scalable amount of agents to easily share information.…”

“…We then use sparse measurements from the robots after canopy closure to estimate the state of the field online. This enables the team of robots to intelligently coordinate their actions, leveraging our earlier work [1], rather than doing a grid search over the whole field. We address a key challenge facing multiagent field-robotic systems operating in spatiotemporally changing domains: managing the tradeoff between learning a robust generalized model and learning a specialized model in order to most accurately explain the current observations of the agents.…”

“…We address a key challenge facing multiagent field-robotic systems operating in spatiotemporally changing domains: managing the tradeoff between learning a robust generalized model and learning a specialized model in order to most accurately explain the current observations of the agents. We use the evolving Gaussian process (E-GP) model [2], trained on ecologically realistic simulations of weed growth [1]. The E-GP aggregates measurements from multiple systems, performs Gaussian process (GP) inference, and then predicts spatiotemporal dynamics throughout a continuous domain using a finite linear dynamical system embedded in a feature space.…”