Verification Report, May 2026

This is the audit trail behind the stack’s pinned values: every build system claim cross-checked against vendor/upstream sources, then confirmed live on the reference device where hardware allows. Each record is VERIFIED, WRONG (+ fix), or UNCLEAR (+ field-confirm action). Read it when you need to know why a pin in versions.env holds its value. Per-unit closeout of the open items lives in FIELD_CONFIRM_RESULTS.md; the symptom-first companion is TROUBLESHOOTING.md.

Method: authoritative URLs (kernel.org, NVIDIA developer docs, Stereolabs, Axelera community, Rock7, MAVLink, MAVROS, ROS, mavlink-router, FastDDS) compared against every magic value, URL, CONFIG_*, command flag, and protocol assumption.

Summary

Category Verified Wrong → fixed Unclear → field-confirm
L4T R36.4.3 / NVIDIA toolchain 4 3 2
Drivers / vendor SDKs 5 6 2
Kernel CONFIG_* names (5.15) 50+ 2 0
Telemetry / MAVLink / MAVROS 8 4 2
Total 65+ 15 6

15 corrections committed across 8 scripts and 4 docs.

Sections: 1 NVIDIA toolchain · 2 drivers and vendor SDKs · 3 kernel CONFIG names · 4 telemetry · 5 May 2026 re-audit · 6 June 2026 re-pin · 7 Phase 8 closeout.

On-device verification (live, confirmed over SSH, 2026-06-10)

The following claims are confirmed on the running board, not just configured. Each command and its observed result is reproducible over SSH.

Claim Confirm command Observed
PREEMPT_RT kernel active uname -v and cat /sys/kernel/realtime uname -v reports SMP PREEMPT_RT (5.15.148-tegra); /sys/kernel/realtime reads 1
Metis NPU on PCIe, bound to driver lspci -d 1f9d: and lspci -k -s 0004:01:00.0 enumerated at 0004:01:00.0 (1f9d:1100), metis.ko loaded
NVMe root findmnt / and lsblk rootfs mounted at / on /dev/nvme0n1p1 (ADATA XPG GAMMIX S55 2 TB)
MAXN_SUPER power mode nvpmodel -q and nproc MAXN_SUPER active (mode ID 0 on this image), 8 cores online, CPU max ~1.98 GHz, EMC locked at max
GPU kernel-side healthy grep CmaTotal /proc/meminfo, ls /sys/class/devfreq/, ls /dev/nvgpu/igpu0, nvgpu lines in dmesg CmaTotal: 262144 kB (device-tree linux,cma pool, no cma= boot arg), GPU devfreq present, GPU at 918 MHz, /dev/nvgpu/igpu0 nodes present, zero nvgpu errors across the boot log
Boot to sshd USB gadget (0955:7020) up, ping 192.168.55.1, then ssh sshd reachable in about 60 seconds; blank HDMI during boot is normal on Orin
GUI desktop session loginctl desktop session on :0 with gdm3 autologin (user j)

Items not covered by the live matrix above, verified live 2026-06-11:

  • ZED X camera: VERIFIED LIVE (2026-06-11). pyzed opens the camera (HD1200@30) and sustains 29.5 FPS stereo with CUDA depth maps after scripts/install_zedx_daemons.sh installs the BMI088/SPSC modules, vendor daemons, and patched libnvisppg.so. See DRIVERS.md §1.5.
  • CUDA userspace: VERIFIED (2026-06-11). /opt/av-env is provisioned (PyTorch cu126, Voyager SDK 1.6.1 wheels), verify_opengl_cuda.sh passes 14/14, and live Metis inference runs at 49.2 FPS end-to-end on 1080p video (29.6 FPS camera-limited on the live ZED X feed). make verify’s venv-import step passes.
  • Sustained throughput beyond these spot checks (the C++ samples, fusion with the --depth-every cadence, GPU load, power, thermals) is measured with the reproducible harness in BENCHMARKS.md; the C++ pipeline itself is documented in ZEDX_METIS_CPP.md.

Section 1, L4T R36.4.3 / NVIDIA toolchain

1.1 ❌ WRONG → FIXED · Board target name

Claim: TARGET_BOARD=jetson-orin-nano-devkit-super. Reality: jetson-orin-nano-devkit-super is the flash config for the Orin Nano Super module (SKU P3767-0004): a physically different module. Using it on an Orin NX 16GB (P3767-0000) installs the wrong power table and misconfigures the SoC. The correct flash target for Orin NX 16GB on the P3768 devkit carrier is jetson-orin-nano-devkit (a symlink to p3768-0000-p3767-0000-a0.conf in the extracted L4T tree; P3509-A02 carriers use the separate p3509-a02-p3767-0000.conf). Fix: versions.env updated; 00_doctor.sh validates against the extracted L4T tree.

Important distinction, NX Super Mode is separate from the -super board target. The Orin NX 16GB gained a MAXN_SUPER nvpmodel profile (uncapped power budget, up to 157 TOPS; the fixed 40 W budget is the separate mode 4) in JetPack 6.2 (L4T R36.4.x+). This is activated at runtime via nvpmodel -m 0, not via the flash config. Live-verified mode table on this image (the flash installs the SUPER nvpmodel conf as the default table): 0=MAXN_SUPER, 1=10W, 2=15W, 3=25W, 4=40W. The flash target remains jetson-orin-nano-devkit for all Orin NX modules regardless of whether Super Mode is used.

Source:

1.2 ❌ WRONG → FIXED · Bootlin toolchain URL

Claim: https://developer.nvidia.com/.../r36_release_v5.0/toolchain/aarch64--glibc--stable-2022.08-1.tar.bz2. Reality: The R36.4.3 page links the toolchain under r36_release_v3.0/toolchain/. The v5.0 URL 404s. Fix: Dockerfile URL corrected. Source: Jetson Linux R36.4.3

1.3 ❌ WRONG → FIXED · GPU devfreq path

Claim: GPU is at /sys/class/devfreq/17000000.ga10b. Reality: R36.4.3 exposes the GPU at /sys/class/devfreq/17000000.gpu/. The .ga10b suffix is R35-era. We now try .gpu first and fall back to .ga10b for backwards compat. Fix: jetson_rt_tune.sh updated to probe .gpu → .ga10b → platform-direct path. Source: R36.4.3 Clocks

1.4 ✅ VERIFIED · Flash command structure

l4t_initrd_flash.sh --external-device nvme0n1p1 -c flash_l4t_t234_nvme.xml -p "-c flash_t234_qspi.xml" --showlogs --network usb0 <board> internal is the canonical R36.4.3 NVMe pattern. Both XMLs ship in R36.4.3. Source: R36.4.3 Flashing Support

1.5 ✅ VERIFIED · Bootlin toolchain version

aarch64--glibc--stable-2022.08-1 (gcc 11.3) is the recommended toolchain for L4T R36.4.3.

1.6 ✅ VERIFIED · CUDA SM 8.7 for Orin NX

Orin NX is GA10B / Ampere, compute capability 8.7. CUDA_ARCH_BIN=8.7 in build_opencv_cuda.sh is correct. Source: CUDA GPUs

1.7 ✅ VERIFIED · EMC clock path

/sys/kernel/debug/bpmp/debug/clk/emc/ with rate, min_rate, max_rate, mrq_rate_locked: correct for R36.4.3.

1.8 ✅ VERIFIED · P3509-class carrier board target

P3509-class carriers use the stock NVIDIA jetson-orin-nano-devkit target with flash_t234_qspi.xml.

1.9 ❓ UNCLEAR · LINK_WAIT_MAX_RETRIES default

Path drivers/pci/controller/dwc/pcie-designware.h is correct. Mainline upstream sets it to 10. NVIDIA’s L4T tree may patch this differently. Field check: grep LINK_WAIT_MAX_RETRIES Linux_for_Tegra/source/kernel/kernel-jammy-src/drivers/pci/controller/dwc/pcie-designware.h after Phase 1 extract; confirm pre-patch value before our sed fires.

1.10 ❓ UNCLEAR · Carrier HV rail and 40 W Super Mode capability

Situation: Orin NX 16GB MAXN_SUPER (the uncapped super profile, up to 157 TOPS; the fixed 40 W budget is the separate mode 4) is available in JetPack 6.2 / L4T R36.4.x+. Activated via nvpmodel -m 0 on this image (live-verified table: 0=MAXN_SUPER, 1=10W, 2=15W, 3=25W, 4=40W). MAXN_SUPER is confirmed active on the bench (2026-06-10); sustained 40 W under flight load remains the open field-confirm item. Gap: NVIDIA’s power documentation states NX Super Mode requires the carrier to expose the HV DC-in rail and be rated for sustained 40 W + thermal dissipation. Verify this against your carrier’s power spec and schematic before enabling MAXN_SUPER. Using it on an under-rated carrier risks brownout or thermal shutdown in flight. Field check before enabling Super Mode:

  1. Confirm carrier input voltage range and DC-DC converter current rating from the carrier vendor.
  2. With nvpmodel -m 0 (MAXN_SUPER) active and Metis inference running, measure rail voltage and monitor tegrastats for throttling (@ suffix) and junction temps.
  3. Keep a thermal margin: if temps exceed 85 °C at idle under MAXN_SUPER, the thermal solution is insufficient. Default: rt-tune defaults to NVPMODEL_MODE=0 (MAXN_SUPER), overridable in /etc/jetson-av/power.conf. To reduce power, set NVPMODEL_MODE=3 (25 W) or another lower mode from the table above. Source: NVIDIA platform power and performance

Section 2, Drivers / vendor SDKs

2.1 ❌ WRONG → DOCUMENTED · ZED X kernel driver source repo

Claim: git clone --branch master https://github.com/stereolabs/zed-driver zedx-driver. Reality: No public Stereolabs repo exists for the ZED X / ZED Link kernel driver. Stereolabs distributes signed .deb packages (e.g. stereolabs-zedlink-mono_<ver>-LI-MAX9296-all-L4T36.x_arm64.deb) and does not publish the source. The .deb’s compiled .ko is built against the stock NVIDIA L4T kernel, vermagic will not match our PREEMPT_RT build. Fix: docs/DRIVERS.md opens with a callout explaining the two paths forward (1) get source via Stereolabs business agreement, (2) skip ZED X for now. The in-tree promotion under drivers/media/i2c/zedx/ is preserved as the right architecture once source is available. Source:

2.2 ✅ VERIFIED · ZED Link Mono uses MAX9296

ZED Link Mono = Maxim MAX9296A deserializer; ZED Link Duo / Quad use MAX96712. Confirmed via Stereolabs .deb filenames (-LI-MAX9296- vs -LI-MAX96712-). Our defconfig + Makefile choice was right. Source: Stereolabs driver install guide

2.3 ❌ WRONG → FIXED · ZED SDK installer flags

Claim: silent skip_drivers skip_python skip_cuda skip_tools. Reality: skip_drivers does NOT exist. Documented silent-mode flags are silent, runtime_only, skip_cuda, skip_python, skip_tools, skip_od_module, skip_hub, nvpmodel=0. Fix: install_zed_sdk.sh invocation changed to silent runtime_only skip_python skip_cuda skip_tools skip_od_module skip_hub nvpmodel=0. Source: ZED SDK Jetson install guide

2.4 ✅ VERIFIED · ZED SDK 5.3 + L4T R36.4.3 + CUDA 12.6

ZED SDK 5.3.0 (released Apr 29 2026) lists Jetson builds for both JetPack 6.2 (L4T 36.4) and JetPack 6.2.2 (L4T 36.5), each on CUDA 12.6. The 5.3 pin is therefore valid for the JetPack 6.2 / R36.4.3 baseline, and would also cover R36.5 if the baseline were bumped. ZED_SDK_VERSION stays at 5.3. Source: ZED Releases

2.5 ✅ VERIFIED · pyzed via get_python_api.py

/usr/local/zed/get_python_api.py is still the canonical install path and auto-detects Python/CUDA. There is no PyPI alternative.

2.6 ❌ WRONG → FIXED · Voyager SDK pip extra-index URL

Claim: https://software.axelera.ai/artifactory/axelera-pypi/. Reality: The pip API requires /api/pypi/<repo>/simple suffix. Correct URL: https://software.axelera.ai/artifactory/api/pypi/axelera-pypi/simple. Fix: versions.env and jetson_first_boot.sh updated. Source: Voyager 1.6 release announcement

2.7 ❌ WRONG → FIXED · Axelera PCI vendor:device ID

Claim: lspci -d :1d60:. Reality: Axelera AI vendor ID is 1f9d, Metis device ID 1100. Every lspci -d :1d60: query in our code would silently fail to match. Fix: axelera_brownout_guard.sh, verify_tuning.sh, install_uav_phase7.sh, docs/UAV_RESILIENCE.md, docs/DRIVERS.md, README.md all updated to 1f9d (or 1f9d:1100). Source: Axelera community thread confirming PCI ID

2.8 ❌ WRONG → FIXED · Metis form factor + PCIe spec

Claim: M.2 2230, PCIe Gen4 x2. Reality: M.2 2280 M-key, PCIe Gen3 x4, ~11.55 W avg / 23.1 W peak. Fix: README.md and docs/DRIVERS.md updated. The axelera_brownout_guard.sh 18 W power cap is still appropriate for the verified 23.1 W peak. Source: Axelera Metis M.2 datasheet

2.9 ❓ UNCLEAR · axdevice --set-power-limit flag spelling

axdevice is the documented Voyager CLI; setting a power limit IS a documented capability. Confirmed 2026-06-11 (FIELD_CONFIRM 3.2): axdevice --help on the installed 1.6.1 runtime lists --set-power-limit LIMIT verbatim. axelera_brownout_guard.sh is correct as written.

2.10 ✅ RESOLVED · AXELERA_GST_EXPLICIT_PARSE=1 was stale, removed

Closed 2026-06-11 (FIELD_CONFIRM 3.3): grep -rn over the full Voyager 1.6.1 checkout and installed wheels returns zero references. The variable does nothing. The export was removed from jetson_first_boot.sh. Inference verified working without it; the decode workaround Jetson actually needs is GST_PLUGIN_FEATURE_RANK=nvv4l2decoder:NONE (AV_STACK.md).

2.11 ✅ VERIFIED · Voyager SDK 1.6 ships pip wheels

axelera-rt and axelera-devkit packages are the current Voyager 1.6 install path; legacy install.sh --driver is deprecated.

Section 3, Kernel CONFIG_* names (Linux 5.15 / L4T R36.4.3)

Verified against torvalds/linux at tag v5.15 via raw.githubusercontent.com. All ~50 flags in our defconfig block were confirmed to exist with the exact names we use EXCEPT:

3.1 ❌ WRONG → FIXED · CONFIG_TPM_HW_RANDOM → CONFIG_HW_RANDOM_TPM

Reality: The symbol is HW_RANDOM_TPM (in drivers/char/hw_random/Kconfig), not TPM_HW_RANDOM. Kconfig silently ignores the wrong name, we got no random source from the TPM with the previous defconfig. Fix: 01_extract_and_patch.sh defconfig block.

3.2 ❌ WRONG → DROPPED · CONFIG_DEVKMEM

Reality: Removed from upstream Linux in 5.13 (commit f2ad42f6db20). Symbol does not exist in 5.15. Our # CONFIG_DEVKMEM is not set was a no-op. Fix: Comment retained but annotated; symbol no longer referenced as if it were live.

3.3 ✅ VERIFIED · CONFIG_USB_ACM

Our addition was correct as a kernel symbol (lives in drivers/usb/class/Kconfig, module cdc-acm). However, see §4.4 for the misconception about what hardware actually needs it.

3.4 ✅ VERIFIED · CONFIG_CPU_ISOLATION

Real bool symbol in init/Kconfig (default y); independent of NO_HZ_FULL. Our usage was right.

3.5 ⚠️ FUTURE WARNING · CONFIG_MEMCG_SWAP

Removed in 6.1 (knob became boot/sysctl). Fine for 5.15 / L4T R36.4.3; flag for any future kernel rebase.

All other 50+ flags verified correct: PREEMPT_RT, NO_HZ_FULL, HZ_1000, RCU_NOCB_CPU, IRQ_FORCED_THREADING, HIGH_RES_TIMERS, RCU_BOOST, PREEMPT_DYNAMIC, NUMA_BALANCING, SCHED_AUTOGROUP, LATENCYTOP, DMABUF_HEAPS{,_CMA,_SYSTEM}, DMABUF_SYSFS_STATS, SYNC_FILE, SW_SYNC, CMA_SIZE_MBYTES, TRANSPARENT_HUGEPAGE{,_MADVISE}, USERFAULTFD, PAGE_REPORTING, CGROUPS, CGROUP_SCHED, CGROUP_CPUACCT, CPUSETS, CGROUP_DEVICE, CGROUP_FREEZER, CGROUP_PIDS, CGROUP_BPF, MEMCG, TCP_CONG_BBR, DEFAULT_TCP_CONG, NET_FOU, BPF_JIT{,_ALWAYS_ON}, XDP_SOCKETS, NET_RX_BUSY_POLL, NET_SCH_FQ{,_CODEL}, BLK_MQ_PCI, NVME_MULTIPATH, NVME_HWMON, IO_URING, USB_ANNOUNCE_NEW_DEVICES, USB_SERIAL_FTDI_SIO, USB_SERIAL_CP210X, USB_USBNET, USB_NET_RNDIS_HOST, USB_NET_CDCETHER, KEXEC{,_FILE}, CRASH_DUMP, PROC_VMCORE, SECURITY{,_YAMA,_LOCKDOWN_LSM}, LSM, TCG_TPM, TCG_TIS, MODVERSIONS, MODULE_SRCVERSION_ALL, MODULE_FORCE_LOAD, HARDENED_USERCOPY, FORTIFY_SOURCE, STACKPROTECTOR_STRONG, RANDOMIZE_BASE, RANDOMIZE_MODULE_REGION_FULL, INIT_STACK_ALL_ZERO, DEVMEM, LEGACY_PTYS, RTW88{,_8822CE}, ARM64_PTR_AUTH, ARM64_BTI{,_KERNEL}, CRYPTO_AES_ARM64_CE, CRYPTO_SHA512_ARM64, KERNEL_MODE_NEON.

Source: https://raw.githubusercontent.com/torvalds/linux/v5.15/<path> for each Kconfig file referenced.

Section 4, Telemetry / MAVLink / MAVROS / Iridium

4.1 ❌ WRONG → FIXED · Pixhawk 6X TELEM2 → /dev/ttyTHS1 (not ttyTHS0)

Claim: TELEM2 maps to /dev/ttyTHS0. Reality: On P3509-class carriers, Pixhawk TELEM2 is typically wired to UART1, exposed as /dev/ttyTHS1. /dev/ttyTHS0 is the debug console. Verify with dmesg | grep ttyTHS on your specific carrier. Fix: versions.env adds FCU_TTY_DEFAULT=/dev/ttyTHS1, telemetry-failover.conf default updated, install_telemetry_failover.sh helper text updated. Source:

4.2 ❌ WRONG → REWRITTEN · RockBLOCK 9704 enumeration + protocol

Claim: RockBLOCK 9704 enumerates as CDC-ACM (/dev/ttyACM0) and uses the AT command set (AT+SBDWB, AT+SBDIX). Reality: The 9704 uses an FTDI USB-TTL bridge (/dev/ttyUSB0 via ftdi_sio), NOT CDC-ACM. The 9704 protocol is JSPR (JSON-based Serial Protocol for REST), NOT AT commands. AT commands belong to the legacy 9602/9603 modems. Fix: install_telemetry_failover.sh rewritten to be model-aware. New env knob IRIDIUM_MODEL (default 9704) selects the sender:

4.3 ❌ WRONG → FIXED · RC_CHANNELS_RAW → RC_CHANNELS

Both messages exist in MAVLink common dialect, but PX4 / ArduPilot now emit RC_CHANNELS (up to 18 channels). Updated the relay’s TRACKED tuple. Source: MAVLink common.xml

4.4 ⚠️ NOTE · CONFIG_USB_ACM was added for the WRONG reason

We added it thinking RockBLOCK 9704 needed it. The 9704 actually uses FTDI (already covered by CONFIG_USB_SERIAL_FTDI_SIO). CONFIG_USB_ACM=m is still useful, it covers the legacy 9602/9603 if anyone uses one, plus various cellular modems that present as CDC-ACM. Kept in defconfig with corrected justification.

4.5 ❌ WRONG → FIXED · MAVROS install_geographiclib path

Claim: bash /opt/ros/humble/lib/mavros/install_geographiclib_datasets.sh. Reality: Canonical invocation is ros2 run mavros install_geographiclib_datasets.sh: the hardcoded path is not portable. Fix: install_av_stack.sh updated. Source: mavros ros2 README

4.6 ✅ VERIFIED · mavlink-router config syntax

Sections, keys, and modes all correct ([General], [UartEndpoint <name>], [UdpEndpoint <name>], Mode = Server|Normal, Address, Port, Device, Baud, TcpServerPort, Log, LogMode = while-armed, ReportStats = true). Source: mavlink-router examples/config.sample

4.7 ✅ VERIFIED · MAVLink message names + GCS sysid 255

GLOBAL_POSITION_INT, ATTITUDE, SYS_STATUS, RC_CHANNELS all canonical. System ID 255 is the de-facto GCS convention.

4.8 ✅ VERIFIED · MAVROS package names

ros-humble-mavros, ros-humble-mavros-extras correct apt names.

4.9 ✅ VERIFIED · MAVROS FCU URL format

/dev/ttyTHS1:921600 valid. Schemes: serial://, serial-hwfc://, udp://, tcp://, tcp-l://.

4.10 ✅ VERIFIED · nav2_bringup navigation_launch.py

Canonical Nav2 bringup launch on Humble.

4.11 ✅ VERIFIED · ros-humble-rosbag2-storage-mcap

Correct apt name.

4.12 ✅ VERIFIED · Isaac ROS: Jazzy-only for 4.x, Humble pinned to 3.x

Isaac ROS 4.4.0 (released 2026-04-30) requires ROS 2 Jazzy on Ubuntu 24.04: Humble is not supported in the 4.x line. Humble operators must use Isaac ROS 3.x. Apt package coverage for ros-humble-isaac-ros-* was field-confirmed sufficient on the reference device 2026-06-11 (FC-4, FIELD_CONFIRM_RESULTS.md §3.4): install_av_stack.sh installs from the release-3 apt repo, keeping the git clone NVIDIA-ISAAC-ROS/isaac_ros_* + colcon build only as a fallback. Plan Jazzy migration when moving to JetPack 7.x.

Source: https://nvidia-isaac-ros.github.io/releases/index.html

4.13 ✅ VERIFIED with caveat · FastDDS UFW port range 7400:7500/udp

Correct for DDS domain 0. Formula 7400 + 250*domainID + offset + 2*participantID means non-default domains need a wider or different range.

Section 5, May 2026 re-audit (upstream version check)

5.1 ❌ WRONG → FIXED · PyTorch version and index domain

Claim: PYTORCH_VERSION=2.8.0, domain pypi.jetson-ai-lab.io Reality (updated 2026-06-10): the index prunes old wheels; it now serves 2.8.0 through 2.11.0 (2.7.0 is gone) and only the pypi.jetson-ai-lab.io host resolves (.dev is dead at the DNS level). The pin is 2.8.0 (torchvision 0.23.0); device-verified download from the live index. Fix: versions.env, scripts/jetson_first_boot.sh, docs/FLASH.md, README.md. Update (June 2026): Jetson AI Lab has since migrated the index host back to pypi.jetson-ai-lab.io. .io is now primary; .dev is kept only as a first-boot fallback. See §6.2. Source: https://pypi.jetson-ai-lab.io/jp6/cu126/

5.2 ❌ WRONG → FIXED · APX recovery USB ID for Orin NX

Claim: USB_ID_APX=0955:7023 Reality: 0955:7023 is the AGX Orin APX ID. Orin NX (P3767) APX ID is 0955:7323. Fix: versions.env, scripts/04_flash_nvme.sh, docs/FLASH.md, docs/QUICKSTART.md, docs/RUNBOOK.md, docs/TROUBLESHOOTING.md, docs/index.md, docs/AUTOMATION.md.

5.3 ❌ WRONG → FIXED · RockBLOCK 9704: IMT not SBD

Claim: “Iridium SBD”, /dev/ttyACM0 at 19200 baud, CONFIG_USB_ACM=m. Reality: 9704 uses Iridium IMT (Messaging Transport), JSPR JSON over FTDI USB-serial (/dev/ttyUSB0) at 230400 baud. Needs CONFIG_USB_SERIAL_FTDI_SIO=m. SBD/ACM/19200 applies to 9602/9603 only. Fix: docs/TELEMETRY_FAILOVER.md, versions.env comment, docs/index.md, README.md. Source: https://docs.groundcontrol.com/iot/rockblock-9704/intro

5.4 ❌ WRONG → FIXED · TELEMETRY_FAILOVER FCU_TTY example

Claim: FCU_TTY=/dev/ttyTHS0 in the config example. Reality: TELEM2 → /dev/ttyTHS1. ttyTHS0 is the debug console. Fix: docs/TELEMETRY_FAILOVER.md config example.

5.5 ⤳ SUPERSEDED BY §6.1 · baseline pin

The May audit verified JetPack 6.2.2 / L4T R36.5.0 as current for Orin NX. The June re-pin (§6.1) deliberately moves the baseline to JetPack 6.2 / L4T R36.4.3 for a validated Isaac ROS combination. Both run kernel 5.15 and CUDA 12.6; 6.2.2 / R36.5 stays a valid Layer-1 target. (version-lint-ok)

5.6 ✅ VERIFIED · CUDA 12.6, SM 8.7, ZED SDK 5.3, MAVROS 2.14.0, Axelera PCI ID 1f9d:1100, /dev/ttyTHS1 at 921600

All confirmed against upstream docs May 2026.

Section 6, June 2026 re-pin and grounded corrections

A second grounding pass cross-checked every component against current NVIDIA, Stereolabs, Axelera, kernel.org, and PX4 documentation. Findings below.

6.1 🔧 RE-PIN · baseline JetPack 6.2.2 / R36.5 → JetPack 6.2 / R36.4.3

No Isaac ROS release was ever validated on JetPack 6.2.2 / L4T R36.5. JetPack-6 Isaac ROS support tops out at L4T 36.4.x: Isaac ROS 3.2 “Update 1” (Jan 2025) added JetPack 6.2 on Humble, and Isaac ROS 4.x moved to JetPack 7 / Jazzy. JetPack 6.2 maps to L4T 36.4.3 (kernel 5.15, CUDA 12.6.10) and still carries the MAXN_SUPER profile, so the uncapped super ceiling (up to 157 TOPS) is unaffected.

CDN note: R36.4.3 BSP/rootfs/sources live under r36_release_v4.3/. The Bootlin toolchain stays at the release-independent r36_release_v3.0/toolchain/ mirror (v4.3/toolchain 404s). All four URLs verified June 2026. Source: https://nvidia-isaac-ros.github.io/releases/index.html , https://developer.nvidia.com/embedded/jetson-linux-r3643

6.2 🔧 FIXED · PyTorch index host migrated .dev → .io

Jetson AI Lab renamed its PyPI index host from pypi.jetson-ai-lab.dev to pypi.jetson-ai-lab.io. .io is now primary in versions.env; jetson_first_boot.sh falls back to .dev. Supersedes the §5.1 domain note. Source: https://forums.developer.nvidia.com/t/packages-removed-from-pypi-jetson-ai-lab-io-indexes/348376

6.3 🔧 FIXED · prose corrections grounded against vendor docs

Was Corrected to Source
“stock CMA reserves 32 MB” Orin reserves CMA via the device-tree linux,cma pool (256 MiB on this board). Live-verified (2026-06-10): a cmdline cma=2048M FAILED to reserve (“cma: Failed to reserve 2048 MiB”) and also bypassed the DT pool, leaving zero CMA; nvgpu then failed its 64 MB contiguous comptag allocation at GPU poweron (ga10b_cbc_alloc_comptags, “DMA alloc FAILED”), cascading into “Unable to recover GR falcon”, FECS init error, no CUDA, no GPU devfreq, and nvpmodel unable to set any mode. Fix: pass NO cma= boot arg; the DT pool then takes over (verified CmaTotal: 262144 kB, zero nvgpu errors, GPU healthy) on-device boot log + /proc/meminfo
“MAXN 25 W / 100 TOPS” as one figure Two axes: 100 TOPS is sparse-INT8 compute, 25 W is the standard module power budget. Super Mode raises both axes: compute up to 157 TOPS, power to the uncapped MAXN_SUPER budget (the fixed 40 W budget is the separate mode 4). Exact per-mode wattages come from the carrier nvpmodel table, not the docs NVIDIA PlatformPowerAndPerformance
“ZED X is a 4K@30 camera” ZED X is 2× 1920×1200 global-shutter @ 60 fps; 4K@30 is the ZED Link Mono capture card throughput ceiling Stereolabs ZED X datasheet
Metis “in-tree, no DKMS” implied as vendor behavior Axelera ships the driver out-of-tree/DKMS-like; in-tree promotion is a repo PREEMPT_RT hardening choice Axelera Metis-on-Jetson article
LINK_WAIT_MAX_RETRIES patch vendor axl-jetson.patch raises 10 → 50; this repo overrides to the PCIE_LINK_WAIT_MAX_RETRIES pin in versions.env, for Metis link-training reliability Axelera bring-up guide
Metis “~18-20 W” framing Axelera rated typical application power is 3.5-9 W; 18 W is a carrier brownout-budget ceiling, not the device draw Axelera Metis M.2 product page
deserializer “MAX9296” (prose) physical part is MAX9296A; kernel symbol CONFIG_SL_DESER_MAX9296 unchanged Stereolabs ZED Link datasheet
lspci -d 1f9d:1100 match on vendor only (lspci -d 1f9d:) so rev-01 boards (1f9d:11aa) also enumerate upstream pci.ids

Files modified by this verification pass 

versions.env                                ← board name, FCU TTY default,
                                              Voyager pip URL, IRIDIUM_MODEL knob
Dockerfile                                  ← Bootlin URL v5.0 → v3.0
scripts/01_extract_and_patch.sh             ← TPM_HW_RANDOM → HW_RANDOM_TPM,
                                              DEVKMEM annotation
scripts/jetson_rt_tune.sh                   ← GPU devfreq path probe
scripts/jetson_first_boot.sh                ← Voyager pip URL
scripts/axelera_brownout_guard.sh           ← PCI vendor 1d60 → 1f9d
scripts/install_uav_phase7.sh               ← PCI vendor 1d60 → 1f9d
scripts/verify_tuning.sh                    ← PCI vendor 1d60 → 1f9d
scripts/install_zed_sdk.sh                  ← drop skip_drivers; use real flags
scripts/install_av_stack.sh                 ← MAVROS install_geographiclib via ros2 run
scripts/install_telemetry_failover.sh       ← FCU TTY, Iridium 9704 model-aware
                                              relay rewrite (JSPR + AT modes)
docs/DRIVERS.md                             ← form factor, ZED X source caveat
docs/UAV_RESILIENCE.md                      ← PCI vendor
README.md                                   ← form factor + PCI ID
docs/VERIFICATION_REPORT.md                 ← THIS FILE

Note: docs/JETSON_AV_PLATFORM_GUIDE.md was touched by this pass but removed in a later documentation trim, so it no longer exists in the tree.

Field-confirm checklist

Of the six original field-confirm items, FC-2, FC-3, and FC-4 are CLOSED (2026-06-11); FC-1, FC-5, and FC-6 remain open. Section 7 tracks each by its FC-n label.

  • FC-1, LINK_WAIT_MAX_RETRIES stock value: verify in Linux_for_Tegra/source/kernel/kernel-jammy-src/drivers/pci/controller/dwc/pcie-designware.h after the extract phase. If NVIDIA already patched it to a sufficient value, the repo’s override (PCIE_LINK_WAIT_MAX_RETRIES=200 in versions.env, for Metis link-training reliability) is harmless but unnecessary.
  • FC-2, axdevice exact flag spelling: CLOSED 2026-06-11: axdevice --help on the installed Voyager 1.6.1 runtime lists --set-power-limit LIMIT verbatim; axelera_brownout_guard.sh is correct as written (see §2.9).
  • FC-3, AXELERA_GST_EXPLICIT_PARSE=1: ✅ CLOSED 2026-06-11: stale, removed (zero references in Voyager 1.6.1; see §2.10).
  • FC-4, Isaac ROS Humble apt package coverage: CLOSED 2026-06-11: apt coverage confirmed sufficient on the reference device (ros-humble-isaac-ros-nitros, -image-pipeline, -visual-slam, -nvblox, -detectnet all from the release-3 repo); no source build needed. See AV_STACK.md and FIELD_CONFIRM_RESULTS.md §3.4.
  • FC-5, /dev/ttyTHS1 enumeration on your carrier: run dmesg | grep tty after flashing to confirm Pixhawk TELEM2 enumerated where expected.
  • FC-6, carrier HV rail and Super Mode (see §1.10): confirm your carrier’s input voltage and DC-DC rating support 40 W sustained before flying with nvpmodel -m 0 (MAXN_SUPER on this image). MAXN_SUPER is confirmed active on the bench; verify under sustained load with tegrastats. This is safety-critical for flight.

Sources index

NVIDIA:

  • https://developer.nvidia.com/embedded/jetson-linux-r3643
  • https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/index.html
  • https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/SD/FlashingSupport.html
  • https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/HR/JetsonModuleAdaptationAndBringUp/JetsonOrinNxNanoSeries.html
  • https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/SD/PlatformPowerAndPerformance/PlatformPowerAndPerformance.html
  • https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/SD/Clocks.html
  • https://developer.nvidia.com/cuda/gpus
  • https://developer.nvidia.com/downloads/jetson-orin-nx-module-series-data-sheet

Stereolabs:

  • https://www.stereolabs.com/developers/drivers
  • https://www.stereolabs.com/docs/embedded/zed-link/install-the-drivers
  • https://www.stereolabs.com/developers/release
  • https://www.stereolabs.com/docs/development/zed-sdk/jetson
  • https://github.com/orgs/stereolabs/repositories
  • https://github.com/stereolabs/zed-python-api

Axelera:

  • https://github.com/axelera-ai-hub/voyager-sdk
  • https://community.axelera.ai/product-updates/voyager-sdk-new-pipeline-builder-and-more-1313
  • https://community.axelera.ai/metis-pcie-7/axelera-metis-pcie-ai-accelerator-not-recognized-by-lspci-145
  • https://axelera.ai/ai-accelerators/metis-m2-ai-acceleration-card
  • https://axelera.ai/hubfs/Axelera_February2025/pdfs/axelera-ai-m2-ai-edge-accelerator-module.pdf

PX4 companion computer docs: https://docs.px4.io/main/en/companion_computer/

MAVLink / MAVROS / mavlink-router:

  • https://mavlink.io/en/messages/common.html
  • https://mavlink.io/en/services/mavlink_id_assignment.html
  • https://github.com/mavlink/mavros/blob/ros2/mavros/README.md
  • https://github.com/mavlink-router/mavlink-router/blob/master/examples/config.sample

Rock7 / GroundControl (Iridium):

  • https://docs.groundcontrol.com/iot/rockblock-9704/hardware
  • https://docs.groundcontrol.com/iot/rockblock-9704/intro
  • https://github.com/rock7/RockBLOCK-9704
  • https://docs.groundcontrol.com/iot/rockblock/user-manual/at-commands

ROS 2 / Isaac ROS / Nav2 / FastDDS:

  • https://nvidia-isaac-ros.github.io/getting_started/index.html
  • https://github.com/NVIDIA-ISAAC-ROS/isaac_ros_common
  • https://github.com/ros-navigation/navigation2/tree/humble/nav2_bringup
  • https://docs.ros.org/en/humble/p/rosbag2_storage_mcap/
  • https://fast-dds.docs.eprosima.com/en/latest/fastdds/discovery/simple.html

Linux kernel (5.15):

  • https://raw.githubusercontent.com/torvalds/linux/v5.15/

Section 7, Phase 8 closeout (May 2026)

This section marks items from earlier sections as closed once the Phase 8 runbook procedures run on hardware. The “Pending FC-n” labels map to the field-confirm checklist above and to the per-unit result slots in FIELD_CONFIRM_RESULTS.md.

VR item Description Status
§1.1 -super board target paragraph CLOSED: index.md rewritten to separate flash config from the nvpmodel unlock; P-number map and P3509 carrier explanation added
§1.9 LINK_WAIT_MAX_RETRIES pre-patch value Pending FC-1
§1.10 Carrier HV rail and MAXN_SUPER capability Pending FC-6
§2.9 axdevice power-limit flag spelling CLOSED 2026-06-11: --set-power-limit confirmed on the installed Voyager 1.6.1 runtime
§2.10 AXELERA_GST_EXPLICIT_PARSE env var CLOSED 2026-06-11: stale, removed
§3 DMABUF zero-copy kernel CONFIG CLOSED: DMABUF_ZEROCOPY.md documents the kernel bridge (axl_dmabuf.c), tracepoints, GStreamer caps, and the verify_dmabuf_zerocopy.sh invariant checker
§4.1 /dev/ttyTHS1 enumeration on carrier Pending FC-5
§4.12 Isaac ROS Humble apt package coverage CLOSED 2026-06-11: apt coverage confirmed sufficient on the reference device (FIELD_CONFIRM_RESULTS.md §3.4)
§5.2 APX USB ID 0955:7023 to 0955:7323 CLOSED: index.md home page updated
§5.5 cuDNN 8 to cuDNN 9.3 (JetPack 6.2.x) CLOSED: index.md home page updated

The remaining Pending items (FC-1, FC-5, FC-6) follow the procedures in FIELD_CONFIRM_RESULTS.md; FC-2, FC-3, and FC-4 closed 2026-06-11 on the reference device. Run verify_dmabuf_zerocopy.sh for DMABUF. Paste results into the <!-- RESULT: --> slots in that document.