Vibration Isolators
Integrated rubber washers on all camera bracket bolt threads to absorb high-frequency chassis rattle.
Edge ML Vision Pipeline
NVIDIA Jetson Nano edge AI configuration, real-time CNN weed detection benchmarks, camera stabilization, and signal wiring integration.
This document summarizes the edge machine learning (ML) pipeline benchmarks, hardware-software configurations, and chassis electrical integration setup of the Smart Agri Four Legged Bot.
1. Embedded Compute & Vision Configuration
Section titled “1. Embedded Compute & Vision Configuration”The machine learning system is deployed directly at the edge to enable real-time detection without depending on cloud processing.
flowchart LR
RealSense["Intel RealSense D435"] -->|USB 3.0 RGB-D| Jetson["NVIDIA Jetson Nano<br>Ubuntu 18.04 + PyTorch"]
Jetson -->|UART Commands| Arduino["Arduino Due<br>Low-Level Control"]
AmiciProbe["Amici Sense Soil Probe"] -->|UART Serial Telemetry| Jetson
Jetson -->|Logging| CSV["Local Data Log CSV"]
- Central Processor: NVIDIA Jetson Nano running Ubuntu 18.04 LTS (JetPack SDK)
- Deep Learning Backend: PyTorch
- Primary Perception Sensor: Intel RealSense D435 depth camera
- Data Channels: RGB frames fused with Active Infrared Stereo depth maps
- Interface: USB 3.0
- Mounting Alignment: Centered on chassis axis, pitched downward at 10° (optimized from 15° to cover near-field crop zone between the front wheels)
2. ML Inference Pipeline Benchmarks
Section titled “2. ML Inference Pipeline Benchmarks”The weed detection task uses a lightweight Convolutional Neural Network (CNN) trained to identify broad-leaf weed clusters around maize crops.
| Performance Metric | Benchmarked Value | Environment |
|---|---|---|
| Input Image Resolution | 640 × 480 pixels | Outdoor RGB-D |
| Average Inference Latency | 135 ms per frame | Over 5-frame average |
| Static Pipeline Frame Rate | 9.2 FPS | Benchmarked stationary |
| In-Field Traversal Frame Rate | 5.0 FPS | Under continuous motion |
| Detection Accuracy | >88.0% | Broad-leaf weed cluster classification |
| Power Consumption | 9.6 W (Peak) | Jetson Nano + RealSense Camera active |
Bounding Box Detection Outputs
Section titled “Bounding Box Detection Outputs”
3. Vibration & Jitter Mitigation
Section titled “3. Vibration & Jitter Mitigation”To prevent camera shake from causing image blur and ML misclassifications, the robot uses a dual mechanical-software stabilization strategy:
Mechanical Countermeasures
Section titled “Mechanical Countermeasures”Damping Spacers
High-density foam inserts situated underneath the battery pack and Jetson base plate to dampen low-frequency bumps.
Balance Counterweights
Geometrically balanced motor shaft mounts to minimize rotational resonance during locomotion.
Software Countermeasures
Section titled “Software Countermeasures”RealSense Frame Averaging
Averages consecutive depth frames to eliminate high-frequency visual and depth noise.
Confidence Filtering
Computes a 3-frame rolling average before confirming and executing weed bounding box triggers.
Sensor Telemetry Sync
Halts locomotion motors during soil probe insertion to completely eliminate vibration noise during telemetry transmission.
4. Hardware System Integration & Wiring
Section titled “4. Hardware System Integration & Wiring”A structured electrical base plate separates signal lines and motor supplies.
Actuation & Low-Level Control Architecture
Section titled “Actuation & Low-Level Control Architecture”Precise and coordinated joint movements are managed using a dual-controller layout:
- Low-Level Controller: An Arduino Due acts as the dedicated real-time actuator controller (84 MHz clock frequency, 12 hardware PWM outputs, and multiple hardware serial UART interfaces).
- High-Level Interface: Commands are received from the NVIDIA Jetson Nano over a dedicated serial connection running at 115,200 baud. The Arduino Due parses these motion paths and translates them into PWM commands.
- Actuators & Drivers:
- Geared DC Motors: Four planetary geared units for shoulder articulation (driven via dual H-bridge motor drivers).
- Brushless Hub Motors: Two distal direct-drive hub motors (driven via high-current Electronic Speed Controllers or ESCs) provide primary locomotion.
Power Rail Separation
Section titled “Power Rail Separation”To prevent motor current draws from causing brownouts or logic resets on the Jetson Nano, power rails are split:
- Logic Rail: 5V rail stepped down from the main 24V bus via a DC-DC buck converter to supply the Jetson Nano, camera, and sensors.
- Geared Motor Rail: A regulated 12V supply bus managed by the H-bridge drivers for shoulder articulation.
- Locomotion Motor Rail: A direct 24V high-current bus from the battery pack to the ESCs for the hub wheels.
EMI Shielding & Signal Isolation
Section titled “EMI Shielding & Signal Isolation”To prevent electromagnetic interference (EMI) from the high-torque actuators from corrupting the low-voltage sensor lines:
- Perimeter Routing: High-current motor lines are routed strictly along the outer edges of the chassis frame.
- Central Data Channels: Low-voltage signal lines are elevated and routed centrally.
- Ferrite Suppression: Ferrite beads are added to motor cable pairs, and all motor logic wires are shielded and grounded to establish common reference planes.