Dandebot

An all-in-one autonomous lawn-care system.

One robot that handles every part of lawn maintenance: weeding, mowing, edging, aerating and blowing.

autonomous-nav
weed-segmentation
low-level-control
diff-drive

Render of Dandebot with its green status lights

README.md

Why lawn care, and what went into the platform.

## problem statement

Residential and commercial lawn care is highly dependent on manual labor, with recurring costs for mowing, edging, weeding, and cleanup. Labor shortages, rising wages, travel between properties, repetitive physical work, and the need for multiple dedicated tools all increase the cost of maintaining outdoor spaces.

Dandebot explores whether a single autonomous mobile platform can reduce that overhead by performing multiple lawn-care tasks (mowing, weeding, edging, and debris blowing) with minimal human intervention.

NVIDIA JetsonSolidWorksROS 2

## implementation notes

Designed a custom tracked mobile platform and mechanical assemblies in SolidWorks.
Integrated NVIDIA Jetson as the onboard compute platform for perception and autonomy.
Built the software stack around ROS 2 for sensor integration, control, and system communication.
Developed autonomous navigation for unstructured outdoor environments.
Integrated perception for obstacle detection, terrain awareness, and task execution.
Designed the platform around modular, task-specific tooling for mowing, weeding, edging, and debris blowing.
Combined perception, navigation, motion control, and tool actuation into an end-to-end autonomous workflow.
Validated the system through real-world outdoor testing rather than simulation alone.

Mechanical design approach

From concept sketches to SolidWorks assemblies of the chassis, tool heads and linear actuators.

Line sketch of the robot's circular tool deck

Line sketch of the weed extraction auger

CAD render of the tracked chassis with the tool module

CAD close-up of the drive motors and couplings

CAD render of the vertical linear actuator

assembly · gearbox, rails and track modules coming together

Mowing tool module with its blade and linear actuator — **tool** · mowing module

Weeding tool module mounted on a linear rail — **tool** · weeding module

Kinematics design approach

Differential-drive kinematics simulated before hardware: trajectory, wheel speeds and orientation.

MATLAB figure of the simulated robot path on an x-y grid — **sim** · robot path on the x–y plane

Four plots: robot trajectory, right wheel speed, left wheel speed and robot orientation — **plots** · trajectory, left/right wheel speed, orientation

Control system design

Cascaded position, velocity and current loops with field-oriented motor control.

Motor driver board wired on the bench — **hardware** · motor driver bring-up

Space vector modulation hexagon plot — **svm** · space-vector modulation

dq-frame and abc-frame current waveforms — **dq / abc** · frame transforms

Block diagram of cascaded position, velocity and current controllers with feedback — **loop** · position → velocity → current controllers with feed-forward and feedback

Build & bring-up

Proof of concept, sensor integration, then machining and assembly.

Laptop showing a costmap during proof-of-concept validation — **poc** · validation on a costmap

Integrating sensors on the robot at the lab bench — **sensors** · integration and validation

End mill machining an aluminium plate — **fab** · machining, fabrication and assembly

Dandebot electronics bay with motor drivers and wiring — **electronics** · drivers, power and wiring in the chassis

integration · controllers and actuator in place

Weed perception · DRoC-Net

Finding each dandelion and the point it grows from.

DRoC-Net pipeline: tiled rosette instance segmentation followed by per-instance growth-point regression, with held-out measurements — **pipeline** · stage 1 segments leaf rosettes in overlapping tiles; stage 2 regresses each plant’s growth point

DRoC-Net is two networks. An RF-DETR Seg-Small model, fine-tuned for one class, segments dandelion leaf rosettes in overlapping 640×640 tiles of the camera frame.

A separate DINOv2 ViT-S/14 regressor then looks at each rosette crop and predicts the growth point, the spot the extraction tool needs to hit.

Training labels come from SAM 3 concept prompts, and the fine-tuned model is evaluated against that zero-shot teacher.

Pipeline overviewPDF · 0.9 MB Segmentation experiment kitPython · zip

Interactive explainer Source on GitHub

DRoC-Net interactive explainer: leaf mask and growth point on a lawn tile — **explainer** · nishanthrajkumarr.github.io/DRoC-Net

Complete coverage

Covering every part of a mapped area, then going home.

indoor · coverage run with the live map overlay

The coverage stack goes from the SLAM map to a costmap and a coverage plan, follows it with Regulated Pure Pursuit, switches to DWA around people, revisits missed patches, then docks.

A MATLAB port of the simulation runs on the same map, with a live plot of the mowed area and the trail.

Coverage simulationMATLAB · zip

Indoor testing

Target detection, extraction and return-to-home in controlled corridors.

multi-target · two targets and return home

single-target · immediate extraction after detection

Outdoor testing

Mapping a real property, then running the full workflow on grass.

Occupancy map beside a satellite view of the property boundary — **commissioning** · onsite mapping and boundary setup

workflow · end-to-end testing on grass

Dandebot on a lawn beside flowering plants — **field** · on the lawn