Hardware Kit

Jetson / CUDA

ARM CPU + NVIDIA GPU for edge AI and accelerated autonomy. Best for heavy perception or large map processing with CUDA-enabled modules.

CUDA Edge GPU Linux

Compute

ARM CPU + NVIDIA GPU

Memory

4-64 GB unified memory

OS/RTOS

JetPack (Ubuntu-based)

Toolchain

JetPack, nvcc, CUDA Toolkit

Power

5-30W power modes

Typical Boards

Jetson Orin Nano / Orin NX
Jetson Xavier NX
Jetson AGX Orin for max throughput

Toolchain + Build Profile

JetPack-matched CUDA Toolkit + nvcc
Match compute capability to deployed Jetson SKU (Orin/Xavier)
CMake with -DENABLE_CUDA=ON and explicit architecture flags when needed
Nsight for profiling

cmake .. -DTARGET=cuda -DENABLE_CUDA=ON
cmake --build .

Pin-Level + Electrical

3.3V logic on GPIO header
CSI camera interfaces (MIPI CSI-2)
I2C/SPI/UART for sensors and peripherals
Use level shifting for 5V devices

Sensors + Peripherals

Multi-camera arrays via CSI
LiDAR via Ethernet/USB
IMU via I2C/SPI

Comms + Networking

Gigabit Ethernet, Wi-Fi (module dependent)
CAN via add-on or USB adapter

Real-Time + Determinism

GPU workloads are async; manage CPU/GPU synchronization
Use pinned memory for stable transfer rates

Memory + Thermal Constraints

Unified memory is shared by CPU and GPU, so map volume must be budgeted conservatively.
Prolonged perception workloads can trigger thermal throttling without active cooling.
Use nvpmodel plus sustained-clock testing before field deployment.

Recommended Vault Modules

3D Voxel Occupancy + CUDA NDT-lite Scan Matching (CUDA) Costmap Inflation (CUDA) Occupancy Grid Mapping (CUDA) Scan Matching (CUDA) Flight Data Recorder Frontier Exploration

Pitfalls + Mitigations

Thermal throttling under sustained GPU load
CPU/GPU sync stalls if transfers are not batched
Power mode misconfiguration reduces throughput

Field Checklist

Set Jetson power mode (nvpmodel) for workload
Profile GPU kernels and avoid small launches
Verify cooling and sustained clocks