arthurpompeu/lecrop-data

LeCropFollow

Latent Space Planning for Navigation in Unstructured Crop Fields

Felipe Tommaselli¹ · Francisco Affonso² · Arthur Pompeu¹ · Gianluca Capezzuto¹
Arun Narenthiran Sivakumar² · Girish Chowdhary² · Marcelo Becker¹

¹ University of Sao Paulo    ² University of Illinois Urbana-Champaign

IEEE Robotics and Automation Letters, 2026

         

---

Models

Please check the Files and versions for our most up-to-date models. For more information, check: https://github.com/Felipe-Tommaselli/lecropfollow

Data

Navigation datasets from a TerraSentia agricultural robot driving under the canopy, extracted from ROS1 bags. Three sources/controllers:

Config	Episodes	Description
CropFollow++ (cpp)	143	Crop-follow / pure-pursuit (logs path)
CROW (crow)	31	iLQR controller (logs crop_lines, goal, ilqr_time)
LeCropFollow (lecrop)	137	MPPI/RL + vision (logs dist_err, head_err, mppi_*, keypoint)

from datasets import load_dataset
ds = load_dataset("arthurpompeu/lecrop-data", "cpp", split="train")
ep = ds[0]
# ep["rgb"]            -> per-episode video (camera)
# ep["odom_pos_x"], ep["odom_vel_x"], ep["cmd_lin_x"], ...  -> time series

Structure

• Each row = one episode: a stretch where the robot drove through the field until it stopped for a while. Episodes were segmented from the (smoothed) odom speed: "moving" when v > 0.05 m/s; a new episode is cut when it stays stopped for >= 5 s. A single bag can yield several episodes; episodes shorter than 2 s or with < 20 messages were dropped. Bags without odometry become a single episode (the whole bag).
• Each column = one signal (a ROS topic), stored as a list (the time series for that episode). Topics have different rates, so each one has its own time vector *_t (seconds, relative to the episode start) and its own length.
• Videos (rgb, lidar_plot, keypoint_vis_*) are the Video type (MP4/H.264): one video per episode at ~10 fps, with per-frame times in *_t. The HF viewer renders a player.
• Removed: rosbag-level fields (header, seq, stamp, frame_id, covariance, layout), heavy raw sensors (depth and LiDAR/PointCloud) and plumbing (tf, camera_info).

Note — browsable version. This published version is downscaled for the dataset viewer: videos are re-encoded to 320×180 and each numeric signal is sub-sampled to <= 250 samples per episode. This keeps trends/shapes intact and makes the viewer fast, but it is not full resolution. For training, request the full-rate / 640×360 variant.

Columns

Metadata (scalars): source, episode, bag, duration_s, n_msgs.

Signals (lists; <g>_t = relative time in s for group <g>):

Group	Columns	Source
odom_*	odom_t, odom_pos_{x,y,z}, odom_quat_{x,y,z,w}, odom_vel_{x,y,z}, odom_angvel_{x,y,z}	/…/dlio/odom_node/odom
imu_*	imu_t, imu_acc_{x,y,z}, imu_gyro_{x,y,z}, imu_quat_{x,y,z,w}	/…/imu
cmd_*	cmd_t, cmd_lin_{x,y,z}, cmd_ang_{x,y,z}	/…/cmd_vel
motion_*	motion_t, motion_lin_{x,y,z}, motion_ang_{x,y,z}	/…/motion_command
path_*	path_t, path_pos_{x,y,z}, path_quat_{x,y,z,w} (list of lists: a polyline per step)	/…/path
goal_*	goal_t, goal_pos_{x,y,z}, goal_quat_{x,y,z,w}	/…/goal (crow)
crop_lines_*	crop_lines_t, crop_lines_{m1,b1,m2,b2} (crop-row lines)	/…/crop_lines (crow)
ilqr_*	ilqr_t, ilqr_time	/…/ilqr_time (crow)
dist_err, head_err	predicted lateral / heading error	/…/*_error_predicted (lecrop)
mppi_dist, elite_scores, value_info	MPPI/RL debug (lists of lists)	/…/rl_debug/* (lecrop)
keypoint	vision keypoints (list of lists)	/…/vision/keypoint (lecrop)

Videos (Video, MP4/H.264, ~10 fps; <g>_t = per-frame time):

Column	Content	Source
rgb	RGB camera video	/…/rgb/image_rect_color/compressed (all)
lidar_plot	LiDAR plot with crop rows	/lidar_plot (crow)
keypoint_vis_argmax, keypoint_vis_heatmap	keypoint-visualization videos	/…/vision/keypoint_vis_*/compressed (lecrop)

Source-specific columns are null when an episode does not have them (e.g. some cropfollowpp_lecropfollow bags inside cpp carry lecrop columns).

License

Code is released under the MIT License. The paper is published under CC BY.