Edge compute — Jetson Nano vs Orin Nano vs RPi 5 (research)

Context

The onboard companion PC must:

• Run the ROS 2 Jazzy stack (sensor_monitor, sweep_node, MAVROS bridge)
• Real-time inference of a trained RL policy (PPO actor, ~1 MB params)
• Optionally — SLAM (slam_toolbox) + Nav2

Comparison table (2026)

Final recommendation — Jetson Orin Nano (8 GB, Super)

Why:

1. 40-67 TOPS — for policy size <10 MB and 100 Hz input, plenty for ~1000 Hz inference (huge margin).
2. CUDA + cuDNN + TensorRT — the RL policy converts to TensorRT for a 3-5× speedup.
3. JetPack 6.0 — Ubuntu 22.04 + ROS 2 Humble apt-installable; Jazzy via source build (~1 hour).
4. Power 7-15 W — on a 3S 5000 mAh BMS gives 25-40 min flight (motors included).
5. 8 GB RAM — enough for ROS 2 + Nav2 + SLAM + policy + parsers simultaneously.

Alternative: RPi 5 (8 GB) — if RL inference isn’t needed (or runs on a separate embedded MCU):

• 6× cheaper
• ARM CPU is enough for ROS 2 + sensor pipeline + MAVROS bridge
• No CUDA → policy inference on CPU (ONNX runtime / NumPy) — realistically 5-20 Hz for a small actor
• Suitable as an “MVP onboard” if RL is offloaded to a ground station

Don’t buy

• Jetson Nano (legacy) — EOL since 2024, Foxy only, Maxwell GPU is weak for modern policies. NVIDIA recommends Orin Nano as the replacement.
• Coral TPU Edge — Google EdgeTPU 4 TOPS, but only for TF Lite quantized models. Not for PyTorch/ONNX out of PPO.

Power-budget calculation

3S 5000 mAh BMS = 11.1 V × 5 Ah = 55 Wh / 125 W = ~26 min hover (theoretical max).

ROS 2 Jazzy on Jetson Orin Nano

JetPack 6.0 ships Ubuntu 22.04 + ROS Humble out of the box. Jazzy via source build:

# https://docs.ros.org/en/jazzy/Installation/Alternatives/Ubuntu-Development-Setup.html
mkdir -p ~/ros2_jazzy/src && cd ~/ros2_jazzy
vcs import --input https://raw.githubusercontent.com/ros2/ros2/jazzy/ros2.repos src
rosdep install --from-paths src --ignore-src -y
colcon build --symlink-install

~1 hour on Orin Nano (8 GB RAM is enough).

RL inference on Orin Nano — pipeline

RL policy (PyTorch .pt)
    → ONNX export (torch.onnx.export)
    → TensorRT optimization (trtexec --onnx=policy.onnx --saveEngine=policy.trt)
    → Runtime inference via TensorRT Python API (~0.5 ms / inference)

Sequential: observation → policy → action @ ~1000 Hz (we use 10 Hz mission FSM → 100× headroom).

Sources

1. NVIDIA Jetson Orin Nano Developer Kit datasheet
2. JetPack 6.0 release notes
3. NVIDIA forum — JetPack 6.0 + ROS 2 Jazzy build threads
4. Raspberry Pi 5 specs — https://www.raspberrypi.com/products/raspberry-pi-5/
5. TensorRT 10.0 docs