Robotic crop monitoring in modern horticulture relies on accurate 3D reconstruction of fruits to estimate yield and guide harvesting. Our lab has recently developed a global optimization approach, GO-VMP, which effectively plans robot motions to maximize fruit coverage while minimizing movement cost [1].

The Problem: Real-world deployment introduces a critical challenge: Pose Uncertainty. Even on calibrated rail systems, mechanical vibrations and kinematic errors cause "drift." Current planners assume perfect localization and often generate "chain-like" exploration paths. When pose errors accumulate, the resulting 3D map suffers from "Ghosting" artifacts—where a single fruit is reconstructed as two disjoint clusters—rendering volume estimation and subsequent manipulation inaccurate (see the figure).

The Goal: This thesis aims to develop a Consistency-Aware View Motion Planner. Instead of assuming perfect poses, the goal is to design a planning strategy that is inherently robust to mechanical inaccuracies.

The Approach: The student will extend the current graph-based optimization framework to account for map consistency.

• The core research question is: How can we modify the path planning formulation to ensure the collected data is self-consistent?

• We will investigate methods to introduce geometric redundancy into the view selection process.

• The objective is to find a balance between exploration (seeing new fruits) and exploitation (reinforcing the consistency of the existing map) to eliminate ghosting artifacts, all within a limited time budget.