Hardware Deployment: RealSense Driver on Mango Pi (H616) & Alternative Platforms

[Jayerwang]

Hi @HenryHuYu,

Thank you for this amazing work! I have successfully reproduced the simulation results (Swarm mode) using the provided PyTorch code.

I am now moving to real-world deployment but facing challenges with the Mango Pi (Quad-core A53, 1GB RAM) setup mentioned in the paper.

1. RealSense Driver on H616: I found that the Allwinner H616 chip is not officially supported by the Intel RealSense SDK.

Could you share how you successfully deployed the driver on this specific board? Did you use a specific kernel or a stripped-down version of the library?

2. Platform Alternatives: Given the driver complexity on H616, I am considering switching to NVIDIA Jetson Orin Nano or Raspberry Pi 5 for better support.

Question: Would the potential difference in system latency (compared to your MCU-like optimization on Mango Pi) affect the agile flight performance significantly?

3.Considering I am new to embedded deployment, would you recommend I stick to the Mango Pi (as in the paper) for strict reproduction, or would a Jetson Orin Nano / Raspberry Pi 5 be a smoother path for a beginner? My main concern is driver compatibility and debugging difficulty.

Any advice would be appreciated!

[HenryHuYu]

Any other hardware platform you mentioned(Jetson Orin Nano / Raspberry Pi 5) should work. However, the dynamic of simulation and real drone is different and will affect performance. A System Identification should be done for better results.

[Jayerwang]

Thanks for the heads-up on System Identification!

Actually，I am now leaning towards a strict reproduction using the Mango Pi setup as described in the paper .

Could I ask for some specific details on how your team managed the embedded deployment on such a constrained platform?

1. Environment Setup:

Did you use the official Debian Server (headless) image?

2. RealSense Driver Implementation:

Did you compile librealsense using the RSUSB backend (-DFORCE_RSUSB_BACKEND=true) to bypass the H616 kernel dependencies?

3. Inference Engine & Optimization:

Runtime: Running standard PyTorch on 1GB RAM is extremely heavy. Did you deploy the policy using ONNX Runtime, NCNN, or TFLite?

Quantization: Did you perform int8 quantization to accelerate inference on the Cortex-A53 CPU?

4. Hardware Engineering:

Power Supply: The D435i draws high peak current. Did you power it directly from the Mango Pi's Type-C port, or did you modify the USB cable (cut VCC) to inject external 5V power (BEC) to prevent board brownouts?

Did you use Betaflight or PX4 (MAVLink) for the low-level flight controller?I need to know which communication library to prepare for the UART interface.

5. Development Process:

Did you use the Mango Pi for the entire debugging phase (sensor calibration, loop testing), or was it only used for the final demonstration after debugging on a more powerful PC?

Knowing these details would save me weeks of trial and error. Thanks again!

Any other hardware platform you mentioned(Jetson Orin Nano / Raspberry Pi 5) should work. However, the dynamic of simulation and real drone is different and will affect performance. A System Identification should be done for better results.

[tae22hee]

I'm also waiting for these details. Hope to hear from you soon!

[Kanan99]

I am also having the issues on flashing linux on my Mango MQ Quad cortex A53 1.5GHz 1 gb ram allwinner h616