You are invited to attend the public defence of Ruben Beumer entitled:
‘World Modeling and Decision-Making for Robots in Agriculture’
Time and location
Date: Tuesday, 3 November 2026
Time: 13:30
Location: Building Atlas, room 0.710 / Den Dolech 2, 5612 AZ Eindhoven
Abstract
This thesis investigates world modeling and decision-making for autonomous robotic systems operating in precision agriculture under uncertainty and resource constraints. It contains four main contributions, spanning both theoretical developments and practical robotic applications.
The first contribution considers selective harvesting of delicate crops, such as table grapes, under uncertainty. A method is presented that integrates multi-view mapping and tracking to improve the quality of crop information, together with a recursive decision-making algorithm based on graphs containing reachability dependencies between the products to optimize the harvesting order with respect to both product quality and execution time, including multi-day harvest planning based on the expected quality development.
The second contribution focuses on semantic world modeling for agricultural robots. A framework is developed that combines probabilistic object mapping with graph-based simultaneous localization and mapping (SLAM), enabling the construction of semantically rich maps while maintaining accurate localization without relying solely on GPS. Detected semantic plant properties, such as plant type and size, not only enrich the resulting map but are also used to improve data association. In addition, an interest heuristic indicates where in the field information is most uncertain, which can guide future sensing.The third contribution explores autonomous mechanical weeding using a small, legged robotic platform. A system is developed that leverages the mobility of a quadruped robot equipped with a custom end-effector, enabling precise, herbicide-free weed removal while reducing soil compaction. The proposed approach is supported by an integrated software architecture and evaluated in both indoor and outdoor environments.
The fourth contribution addresses decision-making under costly sensing and actuation within a partially observable Markov decision process (POMDP) framework. The problem of when to sample and actuate is formulated as a stopping time problem, capturing the trade-off between state-dependent costs and intervention frequency. Due to the intractability of optimal solutions, approximate methods based on relaxed dynamic programming and event-triggered control are developed, providing performance guarantees. In addition, a modified relaxed dynamic programming framework is introduced to explicitly bound the complexity, while still providing performance guarantees, designed through value function approximations and optimization via linear matrix inequalities.Overall, the thesis provides a perspective on uncertainty-aware world modeling and decision-making, demonstrating how theoretical methods and robotic system design can be combined to address key challenges in precision agriculture.