Sparse-Reward Reinforcement Learning for Computer Agriculture Autonomous Weeding Under Delayed Crop-Injury Feedback: An Artificial Intelligence Safety-Constrained Regret Evaluation

Autonomous mechanical weeding in computer agriculture must learn from sparse, delayed crop-injury signals, yet the safety cost of exploratory actions is borne by the crop rather than the learner. the present analysis formalises the problem as a constrained regret minimisation over a safety-constrained Markov decision process in which injury feedback arrives with a lag and the injury budget is finite. A Lagrangian safety-constrained regret bound is derived that separates the statistical cost of sparse rewards from the safety cost of delayed feedback, and the bound is evaluated through a replication-based design analysis of the before-after-control-impact estimator under a declared variance structure (between-cluster SD = 6.0, residual SD = 4.5, intraclass correlation = 0.64). Power was estimated across 2000 design cells over cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis indicates that 52 clusters are required for 80% power at a 5 percentage point effect, and that the intraclass correlation of 0.64 dominates the detectable effect at small cluster counts. The contribution is a safety-constrained regret formulation that makes the injury budget an explicit design parameter rather than an emergent property of reward shaping, together with a design-analysis protocol that quantifies how delayed feedback inflates the cluster requirement for evaluating such learners in the field.