IROS 2026 paper: Receding-horizon planner cuts leaf reconstruction error by 10%

Autonomous plant reconstruction is essential for growth monitoring and yield phenotyping, but large-scale field robots face tight travel budgets and limited compute while needing to pick the highest-value viewpoints. Existing next-best-view (NBV) methods often target rigid objects or simple point-cloud coverage, ignoring task-specific leaf geometry and system constraints. To solve this, Arif Ahmed and colleagues developed a receding-horizon NBV planner that uses a novel Centroid-based Information Gain (CIG) function. CIG measures how well observed points are distributed around the centroid of the current point cloud, giving a dense, task-driven utility score that captures leaf surface coverage more effectively than classic visibility-based metrics. The planner evaluates candidate viewpoints, then looks several steps ahead to avoid greedy, short-sighted decisions.

Benchmarked on the public LAST-STRAW dataset — which contains strawberry plant point clouds across multiple growth stages — the method was compared against an attention-driven NBV baseline using visibility-based information gain. The receding-horizon CIG approach consistently reduced surface reconstruction error and improved geometric fidelity, especially when inter-leaf occlusion increased. The team reports up to a 10% improvement in reconstruction accuracy, while still respecting robot travel constraints. Accepted at IEEE/RSJ IROS 2026, this work demonstrates that combining centroid-referenced information gain with look-ahead planning yields practical, high-fidelity plant models — a key step toward scalable automated phenotyping in real agricultural fields.