Skip to main content
Back to timeline
arXivSource publication:

Orchard robots drop the row-spacing prior: FlowRow reaches 87% success over 900 simulated episodes, against 44% for the strongest generative baseline

Synopsis

In orchards and vineyards with unknown row spacing and no prior map, a mobile robot that must cross rows to reach a goal has previously relied on preset turn maneuvers derived from known field geometry, so arbitrary row changes were out of reach. FlowRow treats this as long-horizon goal-directed navigation, conditioning a flow-matching local planner on depth images and a goal expressed relative to the robot to generate spline paths directly. On RowNavBench, built in Isaac Sim, FlowRow reaches 87% overall success, while the strongest generative behavior-cloning baseline retrained on the same demonstrations reaches 44%.

Source-provided article image: FlowRow: Learning Navigation Behaviors in Agricultural Environments with Long-Horizon Goal-Conditioned Flow Matching Policies
arXiv

Interpretation

A robot can now move from its current row to a goal up to three rows away without knowing row spacing or row count, handling in-row driving, headland turning and row skipping along the way. Earlier perception-based agricultural navigation mainly handled in-row following, while row transitions used rule-based maneuvers assuming known field geometry and could only enter the adjacent row, without skipping or reaching non-adjacent rows. Over 900 simulated episodes in RowNavBench, FlowRow reaches 87% overall success, clearly above NavDP and SanD-Planner; the strongest generative behavior-cloning baseline retrained on the same agricultural demonstrations reaches 44%.

At each step the planner emits only a short local path, yet a goal beyond the planning horizon steers it, so it does not cut straight through the crops toward the target. A gap exists between short-horizon path prediction and distant goals, which often lie beyond what the depth image can observe. Training places the goal beyond the demonstrated segment with some probability, teaching the model to continue in-row and then turn at the headland; on the longest 150 Traverse episodes of RowNavBench, fixed-horizon supervision yields near-zero success, while bi-horizon supervision substantially raises long-horizon success.

The goal is encoded by range and bearing, and the path by spline control poses carrying both position and heading, which together support near-180-degree headland turns. With position-only control points or Cartesian goals, large heading changes are discontinuous in the angle representation and distant goals are hard to represent. Bearing uses sine-cosine pairs to avoid the angle jump, and range is normalized and clipped; under the same bi-horizon supervision, the Cartesian goal with spline action reaches 0.84 success, and bi-horizon supervision also reduces sensitivity to the choice of action representation.

Perspective

Evaluation covers 900 episodes in AgriGen-generated grape vineyards, lemon orchards and apple orchards in Isaac Sim, plus two real zero-shot deployments in apple and pear orchards with different row spacings; the robot uses a depth camera and pose from RTK-GNSS, with the goal given in the body frame.

In real deployment, an out-of-distribution five-row traverse required brief human intervention because a thin signpost near the row entrance was not reliably perceived by the depth sensor; obstacles occluding the camera view were listed as out-of-distribution test conditions. Sim-to-real generalization is listed as future work.