Flashing the G1's Orin NX from JetPack 5.1.1 to 6.2

Symptom

JetPack 5.1.1 (the G1 EDU's stock platform) caps Python at 3.8 and CUDA at 11.4. That ceiling blocks essentially every modern local-AI tool: Ollama's GPU backends require a newer CUDA than 11.4 provides, and most current STT/TTS/VAD libraries assume Python ≥3.10. Working around it means hand-compiling everything from source against an outdated toolchain — which is exactly what an earlier build era did (llama.cpp compiled from source against CUDA 11.4 / compute_87, functional, but a dead end for anything beyond hand-rolled inference).

There's no fixing this in place. The only way past it is a full platform reflash.

Environment

• Unitree G1 EDU, Orin NX module
• Starting point: JetPack 5.1.1, Python 3.8, CUDA 11.4
• Target: JetPack 6.2 (L4T R36.4.3)
• Robot's original 2TB NVMe SSD (not a blank card — the actual boot drive)

What was actually done

1. Downloaded and integrity-verified two files from Unitree's provided source: g1-nx-j6.2.img.bz2 (system image, ~256GB decompressed) and Jetpack_6.2_nx.tar.bz2 (9.3GB, separate NX module firmware package — not the same thing as the system image, and both are required).
2. Physically disassembled the G1's back panel/handle to access the Orin NX's NVMe SSD.
3. Burned the system image directly to the SSD via a USB NVMe enclosure:

   bzip2 -dc g1-nx-j6.2.img.bz2 | dd of=/dev/sda bs=4M status=progress conv=fsync

   ~846 seconds (~14 minutes) at ~303 MB/s.
4. Ran the NX module firmware update — a separate, required step from the system image flash. Entered APX flashing mode via the PWR+REC button combo, confirmed the device was in that mode via lsusb (should show NVIDIA Corp. APX), then ran sudo ./flash_nx_module.sh. Completed in ~8 minutes, confirmed by: *** The target generic has been flashed successfully. ***
5. Reinserted the freshly-imaged SSD, reassembled the robot, powered on, confirmed it booted from the external nvme0n1p1.
6. Enabled max-performance power mode:

   sudo nvpmodel -m 0
   # reboot
   jetson_clocks --show   # confirm NV Power Mode: MAXN, all 8 CPU cores online

7. Installed required post-flash packages: nvidia-l4t-dla-compiler, libcudla-dev-12-6.
8. Full reassembly, clean power-down.

Verified platform state after flash

Component	Version
OS	Ubuntu 22.04.5 LTS, kernel 5.15.148-tegra aarch64
L4T	R36, Revision 4.3 (= JetPack 6.2)
Python	3.10.12 (up from the 3.8 ceiling on JP5.1.1)
CUDA	12.6, Driver 540.4.0

Gotchas worth flagging explicitly

• SSH host key will change after reimaging — expected, not a sign anything went wrong. Remove the old entry from known_hosts rather than troubleshooting a phantom MITM warning.
• nvcc (the CUDA toolkit compiler) is not installed by default — only the CUDA runtime ships out of the box. If you need to compile anything against CUDA from source again (e.g. ctranslate2, llama.cpp), you'll need to install the toolkit separately.
• Keep a rollback path on hand. The NX module firmware layer, once updated, will only boot JetPack-6.2-compatible images going forward — so if you need to roll back to 5.1.1, you need the old firmware bundle and an old system image backup, not just the image. Confirm both are saved before you start.
• We didn't hit it, but it's worth actively watching for on your own flash: some Jetson reflashes reset the RTC to 1970 on the first boot. Not observed here, but cheap to check.

Why this mattered beyond "newer versions"

This wasn't just a version bump for its own sake. JP5.1.1's ceiling was blocking the entire modern local-AI stack outright — Ollama's GPU backends were flatly incompatible with CUDA 11.4, full stop, which is why an earlier build era rejected Ollama and hand-compiled llama.cpp from source just to get any local LLM inference working. JP6.2 removes that ceiling entirely and is what makes a normal, maintainable dependency stack (Ollama, faster-whisper, Pipecat, a single Python 3.11 venv for both the SDK and the pipeline) possible at all.

Time cost: ~1 session for the physical flash and verification (roughly an hour of actual flashing time plus disassembly/reassembly); the decision to commit to the reflash — confirming the platform ceiling was real and not fixable in place — took considerably longer, across earlier sessions of hitting the CUDA 11.4 wall repeatedly.