STEMbot: A Compliant Robot for Under-Canopy Plant Navigation

Presentation video

STEMbot in five minutes

The complete project video covers the motivation, system design, methods, and experimental evaluation.

Overview

Reaching the places conventional crop robots cannot see

The scalability of organic agriculture is partially limited by the labor costs associated with monitoring for pests. While drones and rovers are well-suited for agricultural monitoring from above or next to plants, many pests live on the underside of leaves or on plant stems, making them detectable only after they have caused significant damage.

STEMbot is a miniature climbing robot system designed for autonomous navigation under plant canopies. It integrates a fully geometric PIN-SLAM pipeline with a semantic OcTree, a manifold-constrained A* planner, and ray-tracing goal specification for branch-aware traversal and plant inspection.

7–33 mm demonstrated stem-diameter range

90° demonstrated branch traversal

50 mm tightest demonstrated curvature radius

4 plants autonomous navigation trials

Contributions

A miniature compliant robot that traverses stems, transitions onto branches, and maintains contact while fully inverted.
A semantic OcTree pipeline combining PIN-SLAM, SAM, and CLIP for geometric reconstruction and semantic odometry.
A manifold-constrained A* planner for branch-aware navigation along plant structures.
Hardware experiments on artificial and live plants evaluating traversal, mapping, and autonomous navigation.

Integrated system

Perception, planning, and control in one climbing platform

RGB-D vision builds a geometric and semantic representation of the plant. A receding-horizon planner searches that representation and dispatches discrete locomotion primitives to the robot.

STEMbot system diagram showing sensing and perception, motion planning, and control subsystems. — System architecture: PIN-SLAM and semantic mapping estimate the traversable manifold, A* computes a path, and closed-loop control executes the selected action primitive.

Hardware

Compliant grip without fully enclosing the stem

Two motor pairs drive concave, high-friction EcoFlex wheels. A spring-loaded four-bar linkage maintains contact across varying stem diameters while allowing branches and protrusions to pass through the robot’s open geometry.

Coordinated wheel motion provides three primitives: longitudinal traversal, circumferential yaw, and pitch adjustment. An Intel RealSense D405 camera—incorporating the D401 depth module—supports stereo depth sensing, while a VL6180X time-of-flight sensor feeds a closed-loop pitch controller.

Four 700:1 sub-micro planetary gearmotors
Compliant EcoFlex 00-45 wheel surfaces
Vertical, yaw, and pitch motion primitives
90 Hz time-of-flight feedback control

Exploded STEMbot hardware diagram and illustrations of vertical, yaw, and pitch motion. — Exploded hardware design and coordinated locomotion modes.

Labeled exploded view of STEMbot components including motors, compliant wheels, camera, and time-of-flight sensor. — Exploded view, hardware component breakdown.

Assembly timelapse condensed to one minute.

Perception

Geometry-first localization with semantic structure

PIN-SLAM registers depth observations without relying on repetitive plant appearance. SAM proposes image masks, CLIP classifies them, and high-confidence depth observations are fused into a probabilistic semantic OcTree.

Perception pipeline with RGB and depth input, SAM masks, CLIP-refined semantic classes, and a registered semantic point cloud. — RGB-D observations become semantic point clouds registered in a globally consistent frame.

Semantic manifold estimation

Traversable stem voxels are indexed for efficient spatial queries. Principal component analysis estimates the local surface normal and longitudinal branch heading used by the planner.

Perception demonstration

Traversable plant structure is separated from surrounding geometry. Our bayesian mapping is robust to false negatives and occasional semantic misclassifications.

Motion planning

A* search constrained to the plant manifold

Each state stores a position, surface normal, and proximal or distal branch heading. Candidate actions are projected back to nearby traversable voxels, checked for collision and orientation consistency, and pruned when a branch transition violates docking geometry.

Diagram defining proximal and distal branch heading vectors for STEMbot. — Branch heading is constrained to proximal or distal directions.

Diagram showing a robot state, naive action, and correction back to the nearest traversable voxel. — Action successors are projected back onto the semantic manifold.

Comparison of invalid and valid branch transitions based on heading geometry. — Branch-switch constraints remove physically infeasible successors.

Visibility-goal diagram with Fibonacci-sphere rays cast from an inspection target toward traversable plant surfaces. — Visibility-constrained goal generation by radial ray casting.

Two goal modalities

State goals move the robot to a specified manifold coordinate. Visibility goals instead identify robot states from which a target point lies inside the camera frustum without stem or leaf occlusion.

Only the first action of each planned path is executed. The system then replans from updated odometry, allowing it to account for wheel slip, stem irregularities, and contact not represented by the model.

Traversal capability

Mechanical limits across diameter, curvature, and junction angle

Bench tests isolate individual geometric variables using common cylindrical objects and 3D-printed PLA fixtures.

STEMbot traversing a 33 millimeter frame, a 7 millimeter pen, a 50 millimeter radius curve, and a 90 degree branch. — Reported traversal extremes from the hardware evaluation.

Stem diameter

Nine supplied trials spanning 8–25 mm.

8 mm

9 mm

10 mm

12 mm

14 mm

16 mm

18 mm

20 mm

25 mm

Branch transitions and curvature

The 120° clip documents a failed transition outside the demonstrated 90° capability.

30° junction

60° junction

90° junction

120° · failed

50 mm radius

75 mm radius

100 mm radius

Autonomous experiments

Mapping and navigation across artificial and live plants

The evaluation uses state goals for Dracaena and Ficus and visibility-constrained goals for Monstera and Olea. Robot and map videos below are complete-run timelapses.

Four experimental plant trials showing specimens, initial and final robot poses, plans, semantic maps, and Chamfer error maps. — Experimental overview from the paper: physical specimens, robot poses, representative plans, semantic overlays, and one-way Chamfer error heatmaps.

01

Artificial specimen · State goal

Dracaena

A static artificial plant used to evaluate autonomous navigation and globally consistent semantic reconstruction.

Robot trial

Map and plan

Semantic reconstruction

Chamfer error

02

Live specimen · Visibility goal

Monstera deliciosa

A visibility-constrained trial on a live plant with low color contrast and complex leaf geometry.

Robot trial

Map and plan

Semantic reconstruction

Chamfer error

03

Live specimen · State goal

Ficus lyrata

A state-goal navigation trial on a tall live specimen used to evaluate mapping consistency over an extended trajectory.

Robot trial

Map and plan

Semantic reconstruction

Chamfer error

04

Artificial specimen · Visibility goal

Olea europaea

A visibility-goal trial requiring the robot to reorient before transitioning between artificial branches.

Robot trial

Map and plan

Semantic reconstruction

Chamfer error

Bar chart comparing one-way Chamfer distance for Dracaena, Monstera, Ficus, and Olea. — One-way Chamfer distance by plant and semantic class.

Mapping accuracy

Globally consistent geometry for closed-loop navigation

3.85 mm mean one-way Chamfer distance across artificial specimens

13.36 mm mean one-way Chamfer distance across live specimens

Higher live-plant error is attributed to organic non-rigidity, biological growth between baseline and trial scans, and a semantic misclassification on the Monstera.

Citation

BibTeX

This entry will be updated when the final proceedings metadata becomes available.

@inproceedings{charlick2026stembot,
  author    = {Charlick, Zachary and Roy Choudhury, Nilay and Ma, Haoyu and Huang, Xiaonan and Berenson, Dmitry},
  title     = {{STEMbot}: A Compliant Robot for Under-Canopy Plant Navigation},
  booktitle = {Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year      = {2026}
}

Acknowledgments and funding: TODO — awaiting author-approved funding language.