Benchmarking Evaluation of Sparse-Reward Reinforcement Learning for Computer Agriculture Robotic Weeding Under Crop-Stage Distribution Shift: An Artificial Intelligence Regret Bound for Herbicide Reduction

Robotic weeding controllers trained with sparse terminal rewards are attractive because they require no dense per-plant annotation, yet their performance is benchmarked almost exclusively on the crop stage at which training data were collected. the present analysis treats crop-stage distribution shift as a formal benchmarking problem and derives a regret bound that converts a sparse-reward reinforcement learning policy's stage-shift sensitivity into an interpretable herbicide-reduction guarantee. The controller is modelled as a finite-horizon Markov decision process whose reward is non-zero only at episode termination, and the shift between training and deployment crop stages is represented as a bounded total-variation perturbation of the transition kernel. Under a declared variance structure with between-cluster standard deviation 6.0, residual standard deviation 4.5 and intraclass correlation 0.64, a replication-based evaluation of the before-after-control-impact estimator was conducted across 2000 design cells spanning cluster counts of 8, 16, 24 and 32 and effect sizes of 3, 5 and 8 percentage points. The analysis characterises the operating characteristics of the design rather than reporting observational data. Power reached conventional levels only at the upper end of the cluster grid, and 54 clusters were required for 80% power at a 5 percentage point effect. The regret bound links stage-shift magnitude, reward sparsity and cluster-level variance to a conservative upper limit on herbicide reduction, and it identifies the conditions under which a benchmark result transfers to a new crop stage.