Skip to content

Companion Computer — Jetson Orin Nano

The companion computer is an NVIDIA Jetson Orin Nano mounted inside the ROV enclosure, running fully headless Ubuntu booted from an NVMe SSD. Its role is strictly limited to three functions: MAVLink relay, video passthrough, and depth recording. No AI inference or heavy compute runs on this node.

Hardware Specifications

Parameter Value
Board NVIDIA Jetson Orin Nano Developer Kit
CPU 6-core Arm Cortex-A78AE
GPU 1024-core NVIDIA Ampere (unused for video)
RAM 8 GB LPDDR5
Boot Drive NVMe SSD (also used for .bag data logging)
OS Ubuntu (headless, no desktop environment)
Network Gigabit Ethernet via tether to topside

Service Architecture

All services are managed and controlled through stream_cli.py, a custom Python CLI.

Service Inventory

Service Binary Function
mavlink-router mavlink-routerd MAVLink relay: Cube Orange+ (serial) ↔ Topside (UDP)
ustreamer ×6 ustreamer MJPEG passthrough for 6 USB cameras

Video Pipeline — ustreamer MJPEG Passthrough

Problem

The Orin Nano lacks a hardware NVENC video encoder. Software-encoding 6 simultaneous camera feeds (e.g., H.264 via ffmpeg) would consume the entire CPU budget and introduce encoding latency.

Solution

All 6 USB cameras are selected to output native MJPEG directly from their onboard ISPs. ustreamer operates in passthrough mode — it reads raw MJPEG frames from the camera's V4L2 device and serves them over HTTP without any transcoding.

Per-Stream Resource Cost

Metric Value
CPU per stream < 1% (no transcode)
Bandwidth per stream [INSERT DETAILS HERE: Measured bandwidth per stream at operating resolution]
Total bandwidth (6 streams) [INSERT DETAILS HERE: Aggregate measured bandwidth]
Latency contribution ~1 frame (capture-to-HTTP, no encode buffer)

ustreamer Configuration

Each camera instance runs with parameters:

ustreamer \
  --device /dev/video{N} \
  --host 0.0.0.0 \
  --port 800{N} \
  --format mjpeg \
  --resolution 640x480 \
  --desired-fps 15
  --allow-origin=*

[INSERT DETAILS HERE: USB camera model/make] Resolution 640x480 by default , configurable per camera at launch via: start cam FPS: 15

Devices are autodiscovered at startup using v4l2-ctl --list-devices.Regular USB cameras use the first /dev/videoX node, Intel RealSense uses the fifth.

mavlink-router bridges the Cube Orange+ flight controller (connected via serial UART) to the topside laptop (connected via Ethernet UDP).

Configuration

[UartEndpoint flight_controller]
Device = /dev/ttyACM0
Baud = 115200

[UdpEndpoint topside]
Mode = Normal
Address = [TOPSIDE_IP]
Port = 14550

[INSERT DETAILS HERE: Actual serial device path, baud rate, topside IP address, any additional MAVLink endpoints]

Design Rationale

Running mavlink-router on the Jetson rather than a dedicated microcontroller provides:

  • Software-configurable routing — add/remove endpoints without hardware changes
  • Message filtering — can selectively forward/drop MAVLink message types
  • Logging — optional MAVLink .tlog capture for post-mission analysis

RealSense Recording — rs-record

The Intel RealSense D435i connects directly to the Jetson via USB 3.0. The rs-record utility captures synchronized RGB + depth frames into .bag files on the NVMe SSD.

  • Storage target: Local NVMe SSD (same drive as OS boot)
  • Retrieval: Post-mission via pull_bag.sh (SCP from topside)
  • Format: ROS .bag — compatible with rosbag play and pyrealsense2 offline processing

Recording resolution: 1920 × 1080 (RGB, 30 FPS); 1280 × 720 (stereo depth, 90 FPS).
Typical file size per minute of recording: ~400 MB – 5 GB, depending on resolution and frame rate.

stream_cli.py — Service Management CLI

A custom Python CLI providing a unified interface to start, stop, and monitor all Jetson camera streams and MAVLink routing.

Usage

# Start all camera streams
start all

# Start a specific camera
start cam 8000

# Start a specific camera with custom resolution
start cam 8000 1280x720

# Stop a specific camera
stop cam 8000

# Restart a specific camera
restart cam 8000

# Start MAVLink routing
start mavlink

# Check status of all services
list

# Re-discover cameras
refresh

# Exit and stop everything
exit

Error handling: if a stream fails to start, status shows error in the table. Ctrl+C triggers clean shutdown of all streams and MAVLink. Camera ports start at 8000 and increment by 1 per camera.

Design Rationale

Wrapping systemctl commands behind a purpose-built CLI reduces operator error during time-critical pool runs. A single command replaces multiple sudo systemctl start/stop/status invocations and enforces correct service startup order.