AGRIBOT Hardware Components and Sensor Integration

AGRIBOT’s physical form is a four-wheel skid-steer chassis designed in SolidWorks 2016 and manufactured at SVNIT Surat with assistance from Team DRISHTI. The chassis houses an Nvidia Jetson Nano as the central compute board, which runs all ROS navigation nodes and CNN inference simultaneously. Navigation sensors — a NEO-M8N GPS module and an MPU-9265 9-DOF IMU/magnetometer — are mounted directly on the chassis and connected to the Jetson Nano, while a Raspberry Pi Camera v2 feeds visual data straight into the crop-weed classification model. The robot’s body (base_link) has a mass of 3.4672 kg; each wheel has a mass of 0.411 kg and uses a 10:1 mechanical reduction drive.

Due to the COVID-19 pandemic, full hardware deployment and field testing were not possible. All navigation algorithms and sensor integrations were validated in Gazebo simulation with realistic noise models. Before deploying AGRIBOT in an actual agricultural field, the GPS waypoint accuracy, heading filter parameters, and CNN inference latency should be re-validated on the physical hardware. Simulation-validated parameters may require field-specific tuning.

Hardware Components

Nvidia Jetson Nano

The onboard compute platform that runs all ROS nodes — including gps_converter, data_manipulation, and autonomus_drive — as well as real-time Bonnet CNN inference on the camera stream. Also handles remote SSH access for monitoring and development. See the Configuring Jetson Nano wiki for setup instructions.

NEO-M8N GPS

u-blox NEO-M8N multi-GNSS receiver providing sensor_msgs/NavSatFix messages. In the ROS stack, these are published on /agribot/fix at 5 Hz. The gps_converter node converts these WGS84 coordinates into local Cartesian XY using an Alvin Mercator projection relative to a fixed origin point (lat: 21.1613331649, lon: 72.7870533933).

MPU-9265 IMU / Magnetometer

InvenSense MPU-9265 9-DOF sensor providing 3-axis accelerometer, 3-axis gyroscope, and 3-axis magnetometer data. The magnetometer heading is used as the primary orientation reference for autonomous navigation. An Arduino sketch (imudata.ino) reads the raw I²C registers and publishes the data as a ROS std_msgs/String message on the imu topic at 10 Hz; the Gazebo simulation replaces this with libhector_gazebo_ros_imu.so at 200 Hz.

Raspberry Pi Camera v2

8 MP Sony IMX219 sensor streaming 640×480 RGB images. In Gazebo, this is simulated by libgazebo_ros_camera.so publishing on /agribot/camera/image_raw at 30 fps. The image stream is passed directly to real-time.py for live Bonnet CNN segmentation, producing per-pixel weed/crop/soil labels overlaid on the video output.

IMU Arduino Interface

The physical MPU-9265 is read by an Arduino microcontroller running imudata.ino, located at agribot_ws/src/imudata/imudata.ino. The sketch initializes the MPU via I²C, reads all 14 accelerometer and gyroscope registers at address 0x68, encodes the values into a compact string format, and publishes them as a ROS topic using ros_lib. The encoding format is:

A{AcX}B{AcY}C{AcZ}D{GyX}E{GyY}F{GyZ}G

Below is the complete sketch from the source:

#include <ros.h>
#include <std_msgs/String.h>
#include <Wire.h>

const int MPU_addr = 0x68;  // I2C address of the MPU-9265
int16_t AcX, AcY, AcZ, Tmp, GyX, GyY, GyZ;

// Set up the ros node and publisher
std_msgs::String imu_msg;
ros::Publisher imu("imu", &imu_msg);
ros::NodeHandle nh;

void setup() {
  nh.initNode();
  nh.advertise(imu);

  Wire.begin();
  Wire.beginTransmission(MPU_addr);
  Wire.write(0x6B);  // PWR_MGMT_1 register
  Wire.write(0);     // set to zero (wakes up the MPU-9265)
  Wire.endTransmission(true);
  Serial.begin(57600);
}

long publisher_timer;

void loop() {
  Wire.beginTransmission(MPU_addr);
  Wire.write(0x3B);  // starting with register 0x3B (ACCEL_XOUT_H)
  Wire.endTransmission(false);
  Wire.requestFrom(MPU_addr, 14, true);  // request a total of 14 registers

  AcX = Wire.read() << 8 | Wire.read();  // 0x3B (ACCEL_XOUT_H) & 0x3C (ACCEL_XOUT_L)
  AcY = Wire.read() << 8 | Wire.read();  // 0x3D (ACCEL_YOUT_H) & 0x3E (ACCEL_YOUT_L)
  AcZ = Wire.read() << 8 | Wire.read();  // 0x3F (ACCEL_ZOUT_H) & 0x40 (ACCEL_ZOUT_L)
  Tmp = Wire.read() << 8 | Wire.read();  // 0x41 (TEMP_OUT_H)   & 0x42 (TEMP_OUT_L)
  GyX = Wire.read() << 8 | Wire.read();  // 0x43 (GYRO_XOUT_H)  & 0x44 (GYRO_XOUT_L)
  GyY = Wire.read() << 8 | Wire.read();  // 0x45 (GYRO_YOUT_H)  & 0x46 (GYRO_YOUT_L)
  GyZ = Wire.read() << 8 | Wire.read();  // 0x47 (GYRO_ZOUT_H)  & 0x48 (GYRO_ZOUT_L)

  String data = "A" + String(AcX) + "B" + String(AcY) + "C" + String(AcZ)
              + "D" + String(GyX) + "E" + String(GyY) + "F" + String(GyZ) + "G";
  Serial.println(data);

  int length = data.indexOf("G") + 2;
  char data_final[length + 1];
  data.toCharArray(data_final, length + 1);

  if (millis() > publisher_timer) {
    imu_msg.data = data_final;
    imu.publish(&imu_msg);
    publisher_timer = millis() + 100;  // publish ten times a second (10 Hz)
    nh.spinOnce();
  }
}

The Arduino sketch publishes raw 16-bit register values as a packed string, not a standard sensor_msgs/Imu message. The data_manipulation.py node on the Jetson Nano side parses this string and converts it into engineering units before applying the Kalman filter. In Gazebo simulation, the IMU plugin publishes proper sensor_msgs/Imu messages directly.

Sensor Simulation in Gazebo

All hardware sensors are replicated in Gazebo using the Hector Gazebo plugin suite, configured in agribot_sensors.urdf.xacro. The simulation parameters are tuned to model real-world sensor behavior as closely as possible.

GPS Plugin

Plugin: libhector_gazebo_ros_gps.so

<plugin name="gps_controller" filename="libhector_gazebo_ros_gps.so">
  <alwayson>true</alwayson>
  <robotNamespace>/agribot</robotNamespace>
  <updateRate>5.0</updateRate>
  <bodyname>gps_link</bodyname>
  <topicname>fix</topicname>
  <velocitytopicname>fix_velocity</velocitytopicname>
  <drift>5.0 5.0 5.0</drift>
  <offset>0 0 0</offset>
  <referenceLatitude>21.1613108102</referenceLatitude>
  <referenceLongitude>72.7869817026</referenceLongitude>
  <gaussiannoise>0.1 0.1 0.1</gaussiannoise>
  <velocitydrift>0 0 0</velocitydrift>
  <velocitygaussiannoise>0.1 0.1 0.1</velocitygaussiannoise>
</plugin>

The reference coordinates correspond to the SVNIT campus location. A positional drift of 5.0 m (XYZ) and Gaussian noise of 0.1 are applied to simulate real GPS inaccuracy. This is why the navigation stack implements a Moving Median filter to smooth noisy GPS-derived headings.

IMU Plugin

Plugin: libhector_gazebo_ros_imu.so

<plugin name="imu" filename="libhector_gazebo_ros_imu.so">
  <serviceName>/imu/calibrate</serviceName>
  <robotNamespace>/agribot</robotNamespace>
  <updateRate>200.0</updateRate>
  <bodyName>imu_link</bodyName>
  <frameId>imu_link</frameId>
  <topicName>imu</topicName>
  <rpyOffset>0 0 0</rpyOffset>
  <xyzOffset>0 0 0</xyzOffset>
  <gaussianNoise>0.00000001</gaussianNoise>
  <accelOffset>0 0 -9.79999999586</accelOffset>
  <accelDrift>0.00000001 0.00000001 0.00000001</accelDrift>
  <accelDriftFrequency>0.00000001 0.00000001 0.00000001</accelDriftFrequency>
  <accelGaussianNoise>0.00000001 0.00000001 0.00000001</accelGaussianNoise>
  <rateDrift>0.0 0.0 0.0</rateDrift>
  <rateDriftFrequency>0.0 0.0 0.0</rateDriftFrequency>
  <rateGaussianNoise>0.0 0.0 0.0</rateGaussianNoise>
  <headingDrift>0.0 0.0 0.0</headingDrift>
  <headingDriftFrequency>0.0 0.0 0.0</headingDriftFrequency>
  <headingGaussianNoise>0.0 0.0 0.0</headingGaussianNoise>
</plugin>

The IMU publishes at 200 Hz with near-zero noise, representing a high-quality sensor in simulation. The gravity offset (-9.8 m/s² on Z) is pre-applied. The imu_link is positioned at xyz="0 0.06 0" relative to base_link.

Magnetometer Plugin

Plugin: libhector_gazebo_ros_magnetic.so

<plugin name="magnetometer" filename="libhector_gazebo_ros_magnetic.so">
  <updateRate>5.0</updateRate>
  <robotNamespace>/agribot</robotNamespace>
  <bodyName>base_link</bodyName>
  <topicName>/magnetic</topicName>
  <magnitude>1.0</magnitude>
  <referenceHeading>0.0</referenceHeading>
  <declination>0.0</declination>
  <offset>0 0 0</offset>
  <drift>0.1 0.1 0.1</drift>
  <driftFrequency>10 10 10</driftFrequency>
  <gaussianNoise>0.1 0.1 0.1</gaussianNoise>
</plugin>

The magnetometer publishes at 5 Hz with Gaussian noise of 0.1 and a drift frequency of 10 Hz. This relatively high noise level is intentional — it motivates the Moving Median + Kalman filter pipeline in data_manipulation.py.

Skid-Steer Drive Plugin

Plugin: libgazebo_ros_skid_steer_drive.so

<plugin name="skid_steer_drive_controller"
        filename="libgazebo_ros_skid_steer_drive.so">
  <updateRate>100.0</updateRate>
  <robotNamespace>/agribot</robotNamespace>
  <commandTopic>/agribot/cmd_vel</commandTopic>
  <odometryTopic>odom</odometryTopic>
  <odometryFrame>odom</odometryFrame>
  <leftFrontJoint>base_link_to_left_wheel_F_joint</leftFrontJoint>
  <rightFrontJoint>base_link_to_right_wheel_F_joint</rightFrontJoint>
  <leftRearJoint>base_link_to_left_wheel_R_joint</leftRearJoint>
  <rightRearJoint>base_link_to_right_wheel_R_joint</rightRearJoint>
  <wheelSeparation>0.65</wheelSeparation>
  <wheelDiameter>0.2</wheelDiameter>
  <robotBaseFrame>dummy_link</robotBaseFrame>
  <torque>20</torque>
  <broadcastTF>True</broadcastTF>
  <covariance_x>0.000100</covariance_x>
  <covariance_y>0.000100</covariance_y>
  <covariance_z>0.000100</covariance_z>
  <covariance_yaw>0.0100</covariance_yaw>
</plugin>

The drive plugin runs at 100 Hz with a wheel separation of 0.65 m and wheel diameter of 0.2 m. Maximum torque is 20 Nm per wheel. Odometry covariance values are configured to reflect realistic wheel slip on soft terrain.

ROS Topic Reference

The following table summarizes every sensor topic published in the AGRIBOT system, covering both the real hardware configuration and the Gazebo simulation:

Sensor	ROS Topic	Message Type	Rate
GPS	`/agribot/fix`	`sensor_msgs/NavSatFix`	5 Hz
IMU	`/agribot/imu`	`sensor_msgs/Imu`	200 Hz
Magnetometer	`/magnetic`	`geometry_msgs/Vector3Stamped`	5 Hz
Camera	`/agribot/camera/image_raw`	`sensor_msgs/Image`	30 fps
Odometry	`/odom`	`nav_msgs/Odometry`	100 Hz

In simulation, the camera topic is /agribot/camera/image_raw (set via <imageTopicName>image_raw</imageTopicName> and <robotNamespace>agribot</robotNamespace> in the Gazebo camera plugin). On physical hardware using the Raspberry Pi Camera v2, you may need to remap the topic from your raspicam_node output to match this namespace.

URDF Link Frame Positions

The sensor links are defined as fixed joints relative to base_link in agribot_body.urdf.xacro:

Link	Parent	Offset (xyz)	Purpose
`imu_link`	`base_link`	`0 0.06 0`	IMU/gyroscope frame
`gps_link`	`base_link`	`0 -0.06 0`	GPS antenna phase center
`magnetometer_link`	`base_link`	`0.3 0 0`	Magnetometer — forward-mounted
`camera_link`	`base_link`	`0 0 0`	Camera — co-located with chassis origin
`dummy_link`	—	—	Root physics frame at ground plane
`base_link`	`dummy_link`	`0 0 0.26`	Chassis center, 0.26 m above ground

Overview

Getting Started

Autonomous Navigation

Crop-Weed Classification

ROS Packages

Hardware Components

Nvidia Jetson Nano

NEO-M8N GPS

MPU-9265 IMU / Magnetometer

Raspberry Pi Camera v2

IMU Arduino Interface

Sensor Simulation in Gazebo

GPS Plugin

IMU Plugin

Magnetometer Plugin

Skid-Steer Drive Plugin

ROS Topic Reference

URDF Link Frame Positions

Build docs developers (and LLMs) love

Overview

Getting Started

Autonomous Navigation

Crop-Weed Classification

ROS Packages

Documentation Index

​Hardware Components

Nvidia Jetson Nano

NEO-M8N GPS

MPU-9265 IMU / Magnetometer

Raspberry Pi Camera v2

​IMU Arduino Interface

​Sensor Simulation in Gazebo

​GPS Plugin

​IMU Plugin

​Magnetometer Plugin

​Skid-Steer Drive Plugin

​ROS Topic Reference

​URDF Link Frame Positions

Build docs developers (and LLMs) love

Hardware Components

IMU Arduino Interface

Sensor Simulation in Gazebo

GPS Plugin

IMU Plugin

Magnetometer Plugin

Skid-Steer Drive Plugin

ROS Topic Reference

URDF Link Frame Positions