JetBot: Low-Cost Open-Source 2-Wheel Robot by NVIDIA

A comprehensive guide for building and programming your NVIDIA Jetson-powered robot.

Project Overview

This project uses the NVIDIA Jetson platform to create an intelligent robot capable of autonomous navigation, object detection, and collision avoidance. It combines edge AI processing with robotics hardware for implementing AI-powered robotics applications at the edge.

Hardware Components

• Computing Platform: NVIDIA Jetson Nano 4GB/2GB (Developer Kit Version)

• Chassis: Waveshare JetBot AI Kit chassis with 3D-printed components

• Motors: 2x TT Gear Motors with 6:1 gear ratio

• Power System:

  • 18650 Lithium Battery Pack (7.4V)

  • Waveshare Motor Driver Hat (4tronix)

• Sensors:

  • Camera: Raspberry Pi V2 Camera (8MP) or IMX219-77 camera module

  • Distance sensors: VL53L0X Time-of-Flight sensor

  • IMU: Optional MPU9250 9-DOF sensor

• Additional Components:

  • WS2812 RGB LED array

  • OLED display (optional)

Assembly Instructions

Step 1: Jetson Setup

1. Flash JetPack 4.6+ to microSD card using NVIDIA SDK Manager

2. Complete initial Ubuntu configuration

3. Install JetBot software:

git clone https://github.com/NVIDIA-AI-IOT/jetbot
cd jetbot
sudo python3 setup.py install

Step 2: Hardware Assembly

1. Mount Jetson Nano to chassis baseplate

2. Connect motors to motor controller using PH2.0 connectors

3. Install camera module using CSI-2 ribbon cable

4. Connect battery to power distribution board

Step 3: Software Configuration

1. Configure camera interface:

from jetbot import Camera
camera = Camera.instance(width=300, height=300)

2. Initialize motor controller:

from jetbot import Robot
robot = Robot()

AI Model Implementation

Used Models

• Collision avoidance: ResNet18 trained on synthetic dataset

• Road following: Dronet-style CNN with regression output

Training Process

1. Collect dataset using Jupyter notebook interface

2. Transfer learning using PyTorch

3. Training workflow:

model = models.resnet18(pretrained=True)
model.fc = nn.Linear(512, 2)

Deployment

1. Convert model to TensorRT format

2. Optimize for Jetson using:

import torch2trt
model_trt = torch2trt(model, [data])

Control Software

from jetbot import Robot
import time

robot = Robot()

# Advanced movement with acceleration
def smooth_move(speed=0.5, duration=1.0):
    robot.left_motor.value = speed
    robot.right_motor.value = speed
    time.sleep(duration)
    robot.stop()

# Object-aware movement
def intelligent_move(obstacle_distance):
    if obstacle_distance > 20:  # cm
        robot.forward(0.4)
    else:
        robot.left(0.3)

Operation Guide

1. Power On:

   • Switch battery to ON position

   • Wait for status LED illumination

2. Connecting:

   • Access via browser at http://jetbot-ip:8888

   • Password: jetbot

3. Autonomous Mode:

   • Run Jupyter notebook for:

   • Live object detection

   • Collision-free navigation

   • Road following

Advanced Features

• Real-time Object Detection using SSD-MobileNet

• Gesture Control using MediaPipe models

• ROS Integration (Melodic/Noetic) for SLAM

• Web-based Remote Control with video streaming

Performance Optimization

• Set Jetson to 10W mode:

sudo nvpmodel -m 0

• Enable GPU-accelerated video decoding

• Use mixed precision quantization

• Implement model pruning with TorchPruner

Troubleshooting

• Camera Not Detected:

  sudo systemctl restart nvargus-daemon

• Motor Stuttering:

  • Check battery voltage (>6.5V)

  • Verify PWM frequency settings

• High Latency:

  camera = Camera.instance( fps=15 ) # Reduce frame rate

Future Enhancements

• Multi-modal fusion (camera + LiDAR)

• ROS 2 Humble integration

• Federated learning capabilities

• 5G connectivity for edge-cloud hybrid processing

Additional Resources from Official Guide

• JetBot Assembly Guide

• Collision Avoidance Tutorial

• ROS Integration Docs