Agricultural robots can now handle messy crop rows better — at least according to new field results from LeCropFollow.
Researchers built LeCropFollow to solve a specific failure mode: under-canopy farm robots losing their way when crops are planted irregularly or rows have gaps. Existing systems convert camera feeds into geometric spatial references, which works fine for tidy rows but discards the contextual information needed to handle ambiguity. LeCropFollow instead feeds a self-supervised semantic heatmap through TD-MPC2, a model-based reinforcement learning planner, and optimizes movement decisions inside a learned latent space rather than a geometric one. Field tests in late-stage corn showed the system matched conventional baselines on clean rows and cut semantic navigation failures by 2.4x in plantation gaps.
The more interesting result is zero-shot transfer: the system trained in simulation and deployed directly in physical corn fields without any fine-tuning. That matters because sim-to-real gaps have historically been one of the harder problems in agricultural robotics — getting a robot to generalize from controlled training data to muddy, variable, real-world fields is where most approaches quietly collapse.
Agricultural automation has attracted significant investment and research attention, but unstructured environments have consistently exposed the limits of geometry-first approaches. LeCropFollow's latent planning angle joins a growing body of work applying representation learning to physical robotics — though field results from a single crop type and season should be treated as a promising data point, not a solved problem.