FlowRow: Learning Navigation Behaviors in Agricultural Environments with Long-Horizon Goal-Conditioned Flow Matching Policies

Utkarsh Bajpai, Cédric Pradalier, Stéphanie AravecchiaarXiv 2026 · arXiv

TL;DR

FlowRow is a novel generative local planner that enables agricultural robots to navigate complex row-crop environments using only egocentric depth and a relative goal. By leveraging flow matching and a "bi-horizon" training strategy, it masters intricate behaviors like row-turning and skipping, achieving an 87% success rate—nearly doubling the performance of existing state-of-the-art baselines.

What Problem Does This Paper Address?

Navigating an orchard or vineyard isn't as simple as driving from point A to point B. These are "semi-structured" environments where crops are arranged in long, narrow rows. A target might be physically close (just a few meters away in the next row), but the robot must often drive to the end of the current row, perform a headland turn, and perhaps skip several rows to reach it.

Most current systems handle this using one of two brittle methods:

  1. Map-based tracking: Relying on high-precision GPS (GNSS) and pre-computed paths, which fail under dense canopies or when the signal drifts.
  2. Rule-based state machines: Using hard-coded logic for "row following" and "turning," which lacks the flexibility to handle arbitrary goal positions or irregular field geometries.

The authors seek a "perception-only" policy that can infer these complex maneuvers directly from what the robot sees, without needing a prior map or manual tuning for every field.

Methods

The core of the system is FlowRow, a goal-conditioned policy based on Flow Matching. Unlike traditional diffusion policies that iteratively denoise a signal, flow matching learns a continuous vector field that transforms random noise into a smooth, feasible trajectory.

Key technical components include:

  • SE(2) Action Space: Instead of just predicting $(x, y)$ coordinates, FlowRow generates B-spline paths in $SE(2)$, meaning it explicitly models both the robot's position and its heading. This is critical for wheeled robots that can't move sideways.
  • Polar Goal Encoding: To handle goals that might be far away, the system encodes the target using normalized range and sine/cosine bearing. This prevents mathematical discontinuities when the goal is behind the robot.
  • Bi-Horizon Supervision: This is the "secret sauce." During training, the model is shown goals both within its immediate planning horizon (e.g., 8 meters) and far beyond it (up to 30 meters). This teaches the policy to follow the "flow" of the rows rather than greedily driving straight into a tree to reach a distant target.
  • Receding Horizon Planning: At each step, the model generates 32 candidate paths. These are ranked using a Euclidean Signed Distance Field (ESDF) to ensure they are both safe (clear of obstacles) and progress toward the goal.
The FlowRow architecture: Depth history and a polar goal are fused via a transformer to condition the flow-matching model, which generates candidate SE(2) paths.

Datasets

The authors introduced RowNavBench, a massive new benchmark for agricultural navigation built in NVIDIA’s Isaac Sim.

  • Scale: 900 episodes across 15 different simulated scenes.
  • Diversity: Includes grape vineyards, lemon orchards, and apple orchards. It varies row lengths, spacing, curvature, and even "distractors" like weeds, rocks, and missing plants.
  • Training Data: Collected 178 demonstrations in "weedless" environments, intentionally keeping the training set simpler than the benchmark to test generalization.
Fig. 3: RowNavBench Scenes. Our RowNavBench benchmark spans 900 point-goal navigation episodes across 15 scenes and 3 crop types (Lemon Orchard, Vineyard, Apple Orchard), with diverse row lengths, spacing, curvature, crop placement, missing crops, and weed/stone density. We illustrate the episodic goal types, isometric and top-down views of representative scenes from the 15 scenes of RowNavBench.

Key Insights and Ideas

  • Decoupling Goal and Action Horizons: By training on distant goals, the model learns that the "correct" path to a goal in another row is to continue driving straight until the row ends. This emergent behavior removes the need for a separate "turning" state machine.
  • Perception over Localization: Because FlowRow plans based on current depth observations, it is remarkably robust to localization errors. While a map-tracking "oracle" fails when GPS lags by just one second, FlowRow’s performance remains stable because it "sees" the trees it needs to avoid in real-time.
  • Generative Multi-modality: Flow matching allows the model to represent multiple valid paths to a goal (e.g., turning left vs. turning right at a headland), which the path selection mechanism then filters for safety.

Key Results

  • Benchmark Performance: FlowRow achieved an 87% success rate across all tasks in RowNavBench. For comparison, the strongest baseline (a version of SanD-Planner) only managed 44%.
  • Complex Maneuvers: In "Traverse" episodes (the hardest tasks involving turning and skipping), FlowRow maintained an 82% success rate, while baselines struggled to exceed 10%.
  • Sim-to-Real: The researchers deployed the simulation-trained model zero-shot onto a Husky A200 robot in a real apple orchard. It successfully navigated rows, performed headland turns, and skipped rows without any fine-tuning on real-world data.

Limitations and Future Work

While impressive, the authors note a few areas for growth:

  • Global Planning: FlowRow is a local planner. For extremely large fields, it would still need to be paired with a high-level topological map to decide which general direction to head.
  • Velocity Control: Currently, the robot moves at a constant velocity. Future iterations could learn to slow down for tight turns or speed up in clear corridors.
  • Dynamic Obstacles: The current benchmark focuses on static environments; handling moving workers or other robots remains a future challenge.

Summary updated

Abstract

Autonomous monitoring and intervention in semi-structured agricultural environments, such as orchards and vineyards, require mobile robots to navigate to locations across diverse field configurations. This entails performing navigation behaviors including in-row driving, row turning and skipping, and inter-row traversal. Recent learning-based approaches have enabled robust in-row navigation from onboard perception, but navigation across rows typically requires a priori knowledge of field geometry to specify and tune row-transition maneuvers, preventing arbitrary transitions between rows. We address this challenge by framing navigation in these environments as a goal-directed navigation task, where a long-horizon goal is expressed relative to the robot and a planner generates local paths based on the goal and the robot's perceptual observations. We present FlowRow, a flow-matching-based local planner which models the distribution of diverse agricultural navigation behaviors as $\mathrm{SE}(2)$ paths using the long-horizon goal and depth perception. To enable systematic evaluation, we introduce RowNavBench, a navigation benchmark for semi-structured agricultural environments developed in Isaac Sim, comprising 900 episodes across grape vineyards, lemon orchards, and apple orchards with varying row lengths, row and plant spacings, plant positions, and weed growth. FlowRow achieves 87% success, compared to 44% for the strongest generative behavior cloning baseline retrained on the same demonstrations. We will release our benchmark, data, and code upon acceptance.
BibTeX
@misc{bajpai2026flowrow,
  author = {Utkarsh Bajpai and Cédric Pradalier and Stéphanie Aravecchia},
  title = {{FlowRow: Learning Navigation Behaviors in Agricultural Environments with Long-Horizon Goal-Conditioned Flow Matching Policies}},
  year = {2026},
  eprint = {2611.21035},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.21035},
}
Figure