# NVIDIA Jetson

> Source: https://aiwiki.ai/wiki/nvidia_jetson
> Updated: 2026-07-31
> Fact-checked: 2026-07-25
> Categories: AI Hardware, NVIDIA
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "NVIDIA Jetson." aiwiki.ai, 31 Jul 2026. https://aiwiki.ai/wiki/nvidia_jetson
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**NVIDIA Jetson** is NVIDIA's family of compact, power-efficient system-on-modules (SoMs) that bring GPU-accelerated computing to machines that cannot lean on a data-center connection: robots, drones, cameras, and other devices that have to sense and decide on their own power budget. First released in 2014, the Jetson line has gone through six architectural generations, and by the mid-2020s its highest-performance modules had become the standard onboard computer for [humanoid robot](https://aiwiki.ai/wiki/humanoid_robot) prototyping. Jetson AGX Orin, launched in 2022 with up to 275 trillion operations per second (TOPS) of AI performance, was the module most humanoid-robot teams reached for first; its 2025 successor, [Jetson Thor](https://aiwiki.ai/wiki/jetson_thor), pushed further with a Blackwell-generation GPU built specifically for running large models on a robot's own body rather than over a wireless link.[1][2][3]

Jetson modules pair an [NVIDIA](https://aiwiki.ai/wiki/nvidia) [GPU](https://aiwiki.ai/wiki/gpu) with an Arm CPU and a shared pool of memory on a single circuit board, running a cut-down version of the same [CUDA](https://aiwiki.ai/wiki/cuda) and [TensorRT](https://aiwiki.ai/wiki/tensorrt) software NVIDIA sells for data-center AI. That software compatibility, as much as any single spec, is why Jetson dominates [robotics](https://aiwiki.ai/wiki/robotics) and [edge AI](https://aiwiki.ai/wiki/edge_ai) compute: a model trained on NVIDIA's cloud GPUs can usually be optimized and redeployed onto a Jetson module with the same toolchain, instead of being rewritten for different silicon.[4]

## Architecture: a module, not a full computer

A Jetson product is not a complete computer on its own. Each module is a small board, roughly credit-card to deck-of-cards sized depending on the tier, carrying a system-on-chip (SoC) that integrates an Arm CPU, an NVIDIA GPU, memory controllers, and, on most generations, dedicated accelerators for video and vision. Manufacturers plug that module into a carrier board, either their own design or the reference carrier board NVIDIA ships in its developer kits, which supplies power regulation, connectors, and I/O. The SoM approach lets a robotics company move from one compute generation to the next, for example from Orin to Thor, without redesigning its entire electronics stack, as long as the new module fits the same connector standard.[4]

Unlike a discrete graphics card, a Jetson module has no separate video memory: the CPU and GPU share one pool of LPDDR memory, which simplifies moving data between perception and control code but also puts a hard ceiling on how large a model the board can hold. That ceiling became a headline spec in its own right once robotics teams started trying to run large vision-language-action models on the robot itself rather than in the cloud, which is part of why Jetson Thor's memory jumped to 128GB.[5]

NVIDIA sells Jetson in three tiers, in ascending order of performance and price: Nano (entry-level), NX (mid-range), and AGX (the flagship tier used in the top module of each generation). Each hardware generation carries a code name rather than a version number: Tegra K1, Tegra X1, Tegra X2, Xavier, Orin, and Thor, the same six architecture families NVIDIA has shipped, in order, since 2014.[6] The Thor generation broke that naming pattern: NVIDIA groups its modules under a "Jetson AGX Thor Series" heading but identifies the individual parts by number, as Jetson T5000, T4000, T3000, and T2000, rather than by Nano, NX, and AGX tier names.[4][43]

## History and generations

Jetson began in 2014 with the Jetson TK1, a development board built around the Tegra K1 chip and aimed at researchers experimenting with [computer vision](https://aiwiki.ai/wiki/computer_vision) and early [deep learning](https://aiwiki.ai/wiki/deep_learning) on a mobile power budget. NVIDIA followed with the TX1 (2015) and TX2 (2017), compact credit-card-sized modules that became popular in university robotics labs and early drones because they were among the smallest boards that could run a real [convolutional neural network](https://aiwiki.ai/wiki/convolutional_neural_network) in real time.[6][7]

The 2018 Jetson AGX Xavier marked a shift from research board to industrial product. Its 512-core [Volta](https://aiwiki.ai/wiki/nvidia_volta) GPU, with dedicated Tensor Cores and two Deep Learning Accelerator (DLA) blocks, delivered 32 TOPS of INT8 inference, aimed at autonomous machines rather than lab benchmarks.[8] A smaller, cheaper Xavier NX followed in 2019 and shipped in volume in 2020, delivering 21 TOPS on a board small enough to fit inside a delivery robot or a drone.[9] NVIDIA also opened Jetson to hobbyists and students with the Jetson Nano developer kit in March 2019, a $99 board built around a stripped-down Maxwell GPU with no Tensor Cores that nonetheless put a real, if modest, [AI accelerator](https://aiwiki.ai/wiki/ai_chip) within reach of a maker budget.[10]

| Module | Year | AI performance (peak) | Power (TDP) | Memory |
|---|---|---|---|---|
| Jetson TK1 | 2014 | ~0.3 TFLOPS (FP32, dense) | 10 W | 2 GB LPDDR3 |
| Jetson TX1 | 2015 | ~1.0 TFLOPS (FP16, dense) | 10 W | 4 GB LPDDR4 |
| Jetson TX2 | 2017 | ~1.3 TFLOPS (FP16, dense) | 7.5-15 W | 8 GB LPDDR4 |
| Jetson Nano | 2019 | 0.47 TFLOPS (FP16, dense) | 5-10 W | 4 GB LPDDR4 |
| Jetson AGX Xavier | 2018 | 32 TOPS (INT8, dense) | 10-30 W (up to ~40 W MAX-N) | 32-64 GB LPDDR4x |
| Jetson Xavier NX | 2019/2020 | 21 TOPS (INT8, dense) | 10-20 W | 8 GB LPDDR4x |
| Jetson Orin Nano (Super) | 2023 (2024 update) | up to 67 TOPS (INT8, sparse) | 7-25 W | 4-8 GB LPDDR5 |
| Jetson Orin NX | 2023 | up to 157 TOPS (INT8, sparse) | 10-40 W | 8-16 GB LPDDR5 |
| Jetson AGX Orin | 2022 | up to 275 TOPS (INT8, sparse) | 15-60 W | 32-64 GB LPDDR5 |
| Jetson AGX Thor (T4000) | 2025 | up to 1,200 TFLOPS (FP4, sparse) | 40-70 W | 64 GB LPDDR5X |
| Jetson AGX Thor (T5000) | 2025 | up to 2,070 TFLOPS (FP4, sparse) | 40-130 W | 128 GB LPDDR5X |

Figures are NVIDIA's own published maximums. Xavier and Orin-generation TOPS are measured as INT8 operations, Thor's headline number is FP4 floating-point operations, and pre-Xavier modules predate NVIDIA's TOPS branding and are shown in dense TFLOPS instead; none of these units convert cleanly into one another.[1][3][4][6]

Jetson AGX Orin, announced at NVIDIA's GTC conference with developer kits shipping from March 22, 2022, reset expectations again. Built on NVIDIA's [Ampere](https://aiwiki.ai/wiki/nvidia_ampere) GPU architecture, the top 64GB configuration delivers up to 275 sparse INT8 TOPS with power configurable between 15 and 60 watts, which NVIDIA described at launch as more than 8 times the processing power of AGX Xavier in a comparable footprint.[11] Smaller Orin NX (originally up to 100 TOPS at 10 to 25W, later raised to 157 TOPS at 10 to 40W) and Orin Nano (originally up to 40 TOPS, 7-15W) modules followed through 2023, filling out the mid-range and entry tiers with the same Ampere architecture and software stack.[1][12][45][47] NVIDIA later unlocked extra performance on the entry-level board through a free firmware update: the December 2024 "Orin Nano Super" configuration raised its ceiling from 40 to 67 TOPS and cut the developer kit price from $499 to $249 without changing the underlying silicon. NVIDIA's product page for the resulting Jetson Orin Nano Super Developer Kit lists its performance as "67 INT8 TOPS" and its memory bandwidth as 102 GB/s, up from 68 GB/s, on the same 8GB of 128-bit LPDDR5.[13][48]

### Super Mode and the sparse-versus-dense basis of Jetson TOPS

The December 2024 uplift reached the production modules a month later. JetPack 6.2, published on January 16, 2025, added higher reference power modes and an uncapped MAXN SUPER mode that raises CPU, GPU, DLA, PVA, and SoC clocks across the Orin NX and Orin Nano families. NVIDIA published the resulting figures in both sparse and dense INT8 terms, which is the most explicit statement it has made about what a Jetson TOPS number actually measures.[45]

| Module | Peak INT8 before Super Mode (sparse / dense) | Peak INT8 in Super Mode (sparse / dense) | DRAM bandwidth before / after |
|---|---|---|---|
| Jetson Orin Nano 4GB | 20 / 10 TOPS | 34 / 17 TOPS | 34 / 51 GB/s |
| Jetson Orin Nano 8GB | 40 / 20 TOPS | 67 / 33 TOPS | 68 / 102 GB/s |
| Jetson Orin NX 8GB | 70 / 35 TOPS | 117 / 58 TOPS | 102 / 102 GB/s |
| Jetson Orin NX 16GB | 100 / 50 TOPS | 157 / 78 TOPS | 102 / 102 GB/s |

Two things follow from that table. The number NVIDIA quotes in marketing is always the sparse one, and the dense figure is exactly half of it in every row; sparsity here is Ampere's fine-grained structured sparsity, which requires two of every four weights to be zero and doubles Tensor Core throughput when a network has been pruned and fine-tuned to that pattern.[46] And on the AGX tier the headline is a sum of two different engines rather than one: NVIDIA's Jetson AGX Orin technical brief attributes up to 170 sparse INT8 TOPS to the 64GB module's Ampere GPU and up to 105 sparse INT8 TOPS to its two NVDLA v2 accelerators, which together produce the 275 TOPS figure.[46] The Orin Nano modules carry no DLA at all, so their 67 and 34 TOPS ratings come entirely from the GPU, while the Orin NX 16GB splits its 157 TOPS into 77 from the GPU and 80 from its two DLAs.[45][1] NVIDIA's Jetson Orin NX datasheet states the pre-Super figures in the same style, "up to 100 (Sparse) INT8 TOPs and 50 (Dense) INT8 TOPs" for the 16GB module and 70 sparse against 35 dense for the 8GB.[47] Anyone comparing a Jetson number against a competitor's should check all three variables: precision, sparsity, and whether fixed-function accelerators are counted.

By the early-to-mid 2020s, independent robotics press described Jetson AGX Orin as the effective standard for humanoid and mobile-robot compute, the module most new entrants reached for by default rather than the exception.[2][14]

## Current module lineup

The table below is NVIDIA's published lineup as of July 31, 2026, with each performance figure carrying the precision NVIDIA states for it. Orin figures are sparse INT8 operations; Thor figures are sparse FP4 floating-point operations. The two scales measure different arithmetic and cannot be converted into one another.[1][4][44][46]

| Module | AI performance (as NVIDIA states it) | GPU | CPU | Memory | Power | Status |
|---|---|---|---|---|---|---|
| Jetson T5000 | 2,070 TFLOPS (FP4, sparse) | 2,560-core Blackwell, fifth-gen Tensor Cores, MIG with 10 TPCs, 1.57 GHz | 14-core Arm Neoverse-V3AE, 2.6 GHz | 128 GB 256-bit LPDDR5X, 273 GB/s | 40-130 W | Production; listed available through August 2035 |
| [Jetson T4000](https://aiwiki.ai/wiki/jetson_t4000) | 1,200 TFLOPS (FP4, sparse) | 1,536-core Blackwell, fifth-gen Tensor Cores, MIG with 6 TPCs, 1.53 GHz | 12-core Arm Neoverse-V3AE, 2.6 GHz | 64 GB 256-bit LPDDR5X, 273 GB/s | 40-70 W | Production; listed available through January 2036 |
| Jetson T3000 | 865 FP4 teraflops | Blackwell; core count not published | 8-core Arm Neoverse | 32 GB LPDDR5X, 273 GB/s | Not published; NVIDIA says roughly half T5000 | Announced July 15, 2026; scheduled Q1 2027 |
| Jetson T2000 | 400 FP4 teraflops | Blackwell; core count not published | Not published | 16 GB | Not published | Announced July 15, 2026; scheduled Q1 2027 |
| Jetson AGX Orin 64GB | 275 TOPS (INT8, sparse) | 2,048-core Ampere, 64 Tensor Cores, 1.3 GHz | 12-core Arm Cortex-A78AE, 2.2 GHz | 64 GB 256-bit LPDDR5, 204.8 GB/s | 15-60 W | Production; through January 2032 |
| Jetson AGX Orin Industrial | 248 TOPS (INT8, sparse) | 2,048-core Ampere, 64 Tensor Cores, 1.2 GHz | 12-core Arm Cortex-A78AE, 2.0 GHz | 64 GB 256-bit LPDDR5 with ECC, 204.8 GB/s | 15-75 W | Production; through July 2033 |
| Jetson AGX Orin 32GB | 241 TOPS (INT8, sparse) | 1,792-core Ampere, 56 Tensor Cores, 1.3 GHz | 8-core Arm Cortex-A78AE, 2.2 GHz | 32 GB 256-bit LPDDR5, 204.8 GB/s | 15-60 W | Production; through January 2032 |
| Jetson Orin NX 16GB | 157 TOPS (INT8, sparse); 78 dense | 1,024-core Ampere, 32 Tensor Cores, 1173 MHz | 8-core Arm Cortex-A78AE, 2.0 GHz | 16 GB 128-bit LPDDR5, 102.4 GB/s | 10-40 W | Production; through January 2032 |
| Jetson Orin NX 8GB | 117 TOPS (INT8, sparse); 58 dense | 1,024-core Ampere, 32 Tensor Cores, 1173 MHz | 6-core Arm Cortex-A78AE, 2.0 GHz | 8 GB 128-bit LPDDR5, 102.4 GB/s | 10-40 W | Production; through January 2032 |
| Jetson Orin Nano 8GB | 67 TOPS (INT8, sparse); 33 dense | 1,024-core Ampere, 32 Tensor Cores, 1020 MHz | 6-core Arm Cortex-A78AE, 1.7 GHz | 8 GB 128-bit LPDDR5, 102 GB/s | 7-25 W | Production; through January 2032 |
| Jetson Orin Nano 4GB | 34 TOPS (INT8, sparse); 17 dense | 512-core Ampere, 16 Tensor Cores, 1020 MHz | 6-core Arm Cortex-A78AE, 1.7 GHz | 4 GB 64-bit LPDDR5, 51 GB/s | 7-25 W | Production; through January 2032 |

Notes on the table. The AGX Orin 32GB figure of 241 TOPS is the MAXN_SUPER ceiling introduced in JetPack 7.2; NVIDIA rated the same module at 200 TOPS before that release.[31][32] NVIDIA's July 2026 announcement of the T3000 and T2000 gives their compute in FP4 teraflops without restating the sparse basis it publishes for T5000 and T4000, and had not released datasheets, power tables, connector pinouts, or prices for either module.[43] The Jetson AGX Orin Developer Kit and the Jetson AGX Thor Developer Kit are development hardware, not production modules, and NVIDIA does not publish lifecycle dates for developer kits at all.[28]

Older modules remain orderable on published end dates rather than being quietly discontinued: Jetson AGX Xavier 32GB, AGX Xavier Industrial, Xavier NX 8GB and 16GB, TX2 NX, and TX2i are all listed through July 2027, and the original Jetson Nano through January 2027. Three parts have already reached end of life: Jetson TX2 4GB/8GB and Jetson AGX Xavier 64GB in January 2025, and Jetson TX1 in January 2021. NVIDIA commits to giving at least eight months of notice before a last shipment.[28]

## Jetson Thor: the humanoid-era tier

Jetson Thor is NVIDIA's sixth Jetson generation and the first designed explicitly around [physical AI](https://aiwiki.ai/wiki/physical_ai) and humanoid robots rather than adapted from an automotive or industrial part. It reached general availability on August 25, 2025. The flagship T5000 module pairs a [Blackwell](https://aiwiki.ai/wiki/nvidia_blackwell)-generation GPU (2,560 CUDA cores and 96 fifth-generation Tensor Cores) with a 14-core Arm Neoverse-V3AE CPU and 128GB of LPDDR5X memory, rated at up to 2,070 FP4 teraflops of AI compute within a 40 to 130 watt envelope; a smaller T4000 variant trims that to 1,200 FP4 teraflops and 64GB of memory in a 40 to 70 watt envelope.[3][5] NVIDIA states both of those figures on a sparse FP4 basis and footnotes the 2,070 number as measured at 130 watts.[44] NVIDIA describes the flagship module as delivering roughly 7.5 times the AI compute and 3.5 times the energy efficiency of AGX Orin at a comparable power draw, with the larger memory pool aimed at letting a 70-billion-parameter-class model run directly on the robot.[5] Developer kits list at $3,499, with production T5000 modules priced from roughly $2,999 at volume, according to NVIDIA and press coverage of the launch.[15]

One architectural change is easy to miss: Thor drops the dedicated Deep Learning Accelerator (DLA) block that shipped on every Jetson from Xavier through Orin, relying instead on the Blackwell GPU's transformer-tuned Tensor Cores to carry the attention-heavy math of modern foundation models.[16][5] That is also why Thor's headline figure is quoted in FP4 teraflops rather than the INT8 TOPS used for Orin and earlier chips; the two numbers describe different arithmetic and are not directly comparable. NVIDIA's own Thor specification table lists a PVA v3 vision accelerator and no DLA row at all, which is the clearest confirmation that the Orin-era practice of adding GPU and DLA throughput together no longer applies.[44]

Thor's early adopters, as named by NVIDIA and reported in trade press, include [Boston Dynamics](https://aiwiki.ai/wiki/boston_dynamics) (Atlas), Agility Robotics (Digit), [Amazon Robotics](https://aiwiki.ai/wiki/amazon_robotics), Caterpillar, [Figure](https://aiwiki.ai/wiki/figure_ai), Hexagon, and Medtronic's surgical robots.[3][17] Boston Dynamics has said its electric Atlas runs on Jetson Thor and is an early adopter of NVIDIA's [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t) foundation models, with its engineers training locomotion and manipulation policies in [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab) before deploying them to the robot.[18] [1X Technologies](https://aiwiki.ai/wiki/1x_technologies) has said Jetson Thor is, in its own assessment, the only commercially available module that meets the onboard compute requirements of its [1X Neo](https://aiwiki.ai/wiki/1x_neo) home robot, which uses it to run large models locally for perception, reasoning, and control rather than depending on a network connection.[19] This article covers Thor's place in the broader Jetson lineup; see the dedicated Jetson Thor article for a full specification breakdown.

On July 15, 2026 NVIDIA extended the line downward with two announced modules. The Jetson T3000 combines a Blackwell GPU, an eight-core Neoverse Arm CPU, 32GB of LPDDR5X at 273GB/s, and 25 GbE connectivity, delivering 865 FP4 teraflops in roughly half the size and power of T5000; the Jetson T2000 offers 400 FP4 teraflops with 16GB of memory for lower-cost systems such as autonomous mobile robots, industrial manipulators, and visual AI agents. NVIDIA also announced a separate IGX T3000 variant with integrated functional safety, and [Cosmos 3](https://aiwiki.ai/wiki/nvidia_cosmos_3) Edge, a 4-billion-parameter world foundation model sized for on-device inference on the new parts. Both Jetson modules were scheduled for availability in the first quarter of 2027, with T3000 emulation on the existing AGX Thor Developer Kit promised for JetPack 7.2.1; NVIDIA said T2000 emulation would follow in a later release.[43] As of July 31, 2026 neither module was shipping and JetPack 7.2.1 had not yet appeared on NVIDIA's JetPack download page, where JetPack 7.2 with Jetson Linux 39.2 remained the current release.[31]

## Software stack

Jetson's advantage has never been the silicon alone. NVIDIA pairs every generation with a matching software stack, and that stack is arguably what keeps robotics teams from switching to cheaper hardware.

**JetPack** is the base SDK: a Linux distribution (Jetson Linux, formerly branded L4T for "Linux for Tegra") bundled with CUDA, cuDNN, and TensorRT, plus board-support packages and power-management tools.[20] JetPack 6 covers the Orin generation, while JetPack 7, rolling out through 2026, extends support to Thor and adds a production-grade Yocto Project build alongside the developer-oriented Ubuntu image, a real-time kernel option, and Blackwell's Multi-Instance GPU (MIG) partitioning for running several isolated workloads on one chip.[21][20]

### JetPack 7.2

JetPack 7.2, released in June 2026, changed the generation split described above. It is the first JetPack 7 release to support the entire Orin product family as well as the Jetson AGX Thor Developer Kit and the T5000 and T4000 modules. Its base release, Jetson Linux 39.2.0, is labeled General Availability by NVIDIA and can be used for production purposes.[31][32] That platform status applies to the board-support package; it does not make every framework wheel or application dependency used on top of it an officially supported combination.

| Layer | JetPack 7.2 version or status |
|---|---|
| Jetson Linux | 39.2.0 GA |
| Linux kernel | 6.8 |
| Root file system | Ubuntu 24.04 |
| CPU architecture | aarch64 |
| CUDA | 13.2.1 |
| cuDNN | 9.20.0 |
| TensorRT | 10.16.2 |
| Supported Jetson families | Orin and Thor |
| VPI (vision programming interface) | 4.1.3 |
| PVA runtime | 2.9.1 |
| DeepStream SDK | 9.1 |
| Holoscan SDK | 3.9.0 |
| Isaac ROS | Coming soon |
| Jetson Platform Services | N/A |
| NVIDIA Container Toolkit | 1.19 (with the ISO image) |

The common Ubuntu 24.04, kernel 6.8, and CUDA 13 foundation brings Orin and Thor onto one software branch, reducing the need to maintain separate JetPack 6 and JetPack 7 images.[31][33] JetPack 7.2 also introduces a unified ISO installer for Orin and Thor developer kits and official OpenEmbedded/Yocto recipes based on the same board-support package. NVIDIA does not provide an SD-card image for the Orin Nano developer kit for this release; its official instructions use the ISO on a USB drive instead.[31][32]

Several features have narrower support than the base release. The new MAXN_SUPER mode raises NVIDIA's stated peak for the 32GB AGX Orin from 200 to 241 TOPS. MIG, by contrast, is a technology preview limited to Thor T5000: the 39.2 release notes say it is not supported on T4000 or any Orin module and initially recommend only two profiles on AGX Thor. NVIDIA's JetPack 7.2 component table also still listed Isaac ROS as "Coming soon" in July 2026, so installing upstream ROS 2 Jazzy does not by itself establish Isaac ROS support for the release.[31][32][33]

The same component table is worth reading closely for two other entries. **NVIDIA DeepStream**, the multi-stream video analytics SDK, is listed at version 9.1 for JetPack 7.2, so the pairing of DeepStream 9.1 with JetPack 7.2 is NVIDIA's own stated combination rather than a community claim.[31] **Jetson Platform Services**, the collection of microservices for generative-AI and vision workflows that shipped as part of JetPack 6, is listed as "N/A" for JetPack 7.2, meaning teams that built on it have no supported JetPack 7 equivalent yet.[31] Alongside those, JetPack 7 aligns Jetson software with the Server Base System Architecture (SBSA) used by Arm server designs, which is what allows Thor to install a unified CUDA 13.0 build shared with server-class Arm targets, and JetPack 7.2 adds native single-command installation of NVIDIA NemoClaw plus a v2.0.0 SIPL camera API package with a unified framework for GMSL and camera-over-Ethernet sensors.[20][31] Jetson Linux 39.2 is dated June 2, 2026 on NVIDIA's download page.[31]

### SO-101 field deployment on Orin

On July 22, 2026, NVIDIA Robotics promoted a setup by Iulia Feroli with the description "from a fresh install to teleoperating SO-101 robot arms in only 30 minutes." The NVIDIA short link leads to an 8-minute-50-second edited video uploaded on July 15 by the Back to Engineering channel, not by NVIDIA. The video description says the series was made with Seeed Studio, and Feroli's companion repository describes itself as field notes rather than official guidance.[34][35][36]

The example uses a Seeed Studio reComputer Super with a Jetson Orin NX 16GB module and J401 carrier. Its recorded software combination is more specific than "JetPack 7.2":

| Layer | Tested value |
|---|---|
| Jetson module | Orin NX 16GB |
| Jetson Linux | r39.2 on Ubuntu 24.04 |
| CUDA | 13.2 |
| Python | 3.12 |
| PyTorch | 2.12.1+cu132 prerelease |
| torchvision | 0.27.1+cu132 prerelease |
| NumPy | 2.2.6 |
| ROS | ROS 2 Jazzy, installed system-wide |
| Robot software | [LeRobot](https://aiwiki.ai/wiki/lerobot) 0.6.0 |
| Manipulators | SO-101 leader and follower arms |

After flashing JetPack, the procedure exposes the JetPack CUDA paths, creates a Python virtual environment, installs the CUDA 13.2 aarch64 PyTorch wheel, and verifies real GPU work with a matrix multiplication instead of checking only whether CUDA is visible. ROS 2 Jazzy is installed at system level, outside that virtual environment. The final steps install LeRobot and its Feetech servo dependencies, identify the two USB serial ports, configure and calibrate the leader and follower arms, and launch local leader-follower teleoperation. Those arm steps follow LeRobot's documented sequence of finding ports, setting up motors, calibrating both devices, and running teleoperation.[36][38]

This example is useful because it exposes the version coupling hidden by JetPack's integrated image, but it is not a supported reference stack. The PyTorch and torchvision packages are prerelease `cu132` wheels. NVIDIA's PyTorch for Jetson compatibility table, updated July 9, listed framework containers only through JetPack 7.1 and no framework wheel for those JetPack 7 entries; JetPack 7.2 was not yet listed.[37] The field notes also report that an automatic Jetson container selection found no r39 image and fell back to a JetPack 6 / r36.4 image that could not run against the JetPack 7.2 host stack.[36]

There is a second dependency exception at the application layer. LeRobot 0.6.0 declared `torch<2.12.0`, while the tested wheel identifies itself as the 2.12.1 prerelease. The guide therefore installs LeRobot with `--no-deps`, adds dependencies manually to avoid replacing the working CUDA wheel, and warns that `pip check` is expected to report version conflicts. It also tests only one Orin NX 16GB system. Other Orin modules share the wheel's `sm_87` target and are expected by the guide to work, but they still require verification; Thor uses `sm_110`, so the Orin wheel is not a Thor installation path.[36]

The result shown is local SO-101 teleoperation. ROS 2 and LeRobot run side by side rather than through a ROS-native LeRobot control bridge, and the walkthrough does not demonstrate policy training, autonomous inference, long-duration stability, or a production deployment. The "30 minutes" figure is the tutorial's clean-path claim, repeated by NVIDIA Robotics, not an independently timed benchmark. The distinction matters: JetPack 7.2 and Jetson Linux 39.2 are production-capable GA platform releases, while this particular PyTorch and LeRobot environment remains a third-party workaround using prerelease packages and an explicit dependency override.[32][34][35][36]


**TensorRT** is NVIDIA's inference optimizer: it takes a trained model exported from PyTorch, TensorFlow, or the ONNX exchange format and compiles it into a runtime engine tuned for a specific Jetson module, using layer fusion and reduced-precision quantization (FP16, INT8, or FP4 on Thor) to cut latency before the model ever runs on the robot.[22] **Isaac ROS** layers robotics-specific, GPU-accelerated packages on top of the open-source ROS 2 framework, covering perception tasks such as stereo depth estimation, object detection, and visual odometry, along with a motion-planning library called Isaac ROS cuMotion.[23] For fixed-camera and infrastructure applications rather than mobile robots, NVIDIA offers a separate platform, **Metropolis**, for vision-AI tasks like retail analytics and factory inspection; it targets the same Jetson hardware but a different use case.[24]

Higher up the stack, [NVIDIA Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim) provides a physics-accurate simulator, built on NVIDIA's Omniverse platform, where robot policies can be tested before they touch real hardware, and Isaac Lab builds on it with a GPU-parallelized framework for training policies with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) at scale, running many simulated robots in parallel to shrink training time and narrow the [sim-to-real transfer](https://aiwiki.ai/wiki/sim_to_real_transfer) gap. Isaac GR00T sits at the top of the stack as NVIDIA's family of open [foundation models](https://aiwiki.ai/wiki/foundation_model) for humanoid manipulation and locomotion, trained in simulation and from [imitation learning](https://aiwiki.ai/wiki/imitation_learning) on human demonstration data, then deployed onto Jetson hardware for on-robot inference.[25] NVIDIA frames the whole chain, Isaac Sim and Isaac Lab for training, Isaac ROS for deployment, and Jetson for on-robot inference, as one connected pipeline rather than a set of separate products.[25]

## On-device generative AI and Jetson agent skills

By 2026 NVIDIA had shifted its Jetson messaging from computer vision toward running open generative models on the module itself. A developer-facing post published on July 28, 2026 by Matthew Leib, built around a video with Sarah Guo of the venture firm Conviction, maps three modules onto three levels of ambition: Jetson Orin Nano Super at 67 TOPS for a first robot, Jetson AGX Orin at 275 TOPS for advanced coursework and applied research, and Jetson AGX Thor at up to 2,070 FP4 teraflops with 128GB of memory for the largest workloads.[39] Those three numbers are accurate against NVIDIA's own module pages, but the post prints them side by side without noting that the two Orin figures are sparse INT8 integer operations while the Thor figure is sparse FP4 floating-point; the Orin Nano Super's 67 TOPS and Thor's 2,070 teraflops are not points on one scale.[1][44][45]

The day before that post, NVIDIA Robotics announced a July 28 livestream, at 9 a.m. Pacific, on running open-source models such as [Gemma 4](https://aiwiki.ai/wiki/gemma_4) and [Qwen](https://aiwiki.ai/wiki/qwen) directly on a Jetson device "with no cloud required", covering how to launch models with [Ollama](https://aiwiki.ai/wiki/ollama), [vLLM](https://aiwiki.ai/wiki/vllm), and [llama.cpp](https://aiwiki.ai/wiki/llama_cpp) and how to pick an inference engine for a given workload.[40] The blog post describes the wider series as three modules covering generative AI, agent building, and vision-language and [vision-language-action](https://aiwiki.ai/wiki/vision_language_action_model) models.[39] Both vLLM and SGLang are listed as supported AI frameworks in the JetPack SDK itself.[20]

NVIDIA used the post to showcase community and university projects, each tied to a specific module:

| Project | Module NVIDIA names | What NVIDIA describes |
|---|---|---|
| SidewalkPilot | Jetson Orin Nano Super | A custom self-driving model executing maneuvers autonomously in a toy electric vehicle |
| Reachy Mini Jetson Assistant | Jetson Orin Nano Super | A low-latency on-device voice and vision assistant for Reachy Mini Lite, running with no cloud, no API keys, and no internet at runtime |
| First AI robot from scratch (Coding with Lewis) | Jetson Orin Nano Super | A robot built around [Mistral](https://aiwiki.ai/wiki/mistral), an open-weight model, by a first-time robotics developer |
| Robotics AI Podcast (Asier Arnaz) | Jetson Orin Nano Super | A Yocto-based video podcast in which two AI models discuss topics in real time |
| Live VLM WebUI | Jetson AGX Orin | A browser interface streaming live [vision-language model](https://aiwiki.ai/wiki/vision_language_model) inference from a camera feed |
| SMoRes | Jetson AGX Orin | A [Carnegie Mellon University](https://aiwiki.ai/wiki/carnegie_mellon_university) robotics team's system that builds a 3D map of an environment while searching for survivors in time-critical rescue scenarios |
| Matcha Bot | Jetson AGX Thor | A first-place hackathon entry from the University of Illinois Urbana-Champaign SIGRobotics team, using two robotic arms running Isaac GR00T N1.5 to pour, prepare, and whisk matcha |
| Multimodal AI Studio | Jetson AGX Thor | A development environment combining vision, audio, and language inputs for multimodal experimentation at the edge |

These are NVIDIA's own descriptions of third-party work; independent performance measurements are not published for any of them.[39]

### Jetson Device Skills and Jetson BSP Skills

The post also points developers at "NVIDIA Jetson Device Skills" and "Jetson BSP Skills" as a way to "harness coding AI agents to create, optimize and deploy real-world AI at the edge."[39] Despite the hardware-sounding names, these are not firmware, a library, or a model. They are two open-source repositories of [agent skills](https://aiwiki.ai/wiki/agent_skills): directories containing a `SKILL.md` instruction file, helper scripts, and reference material, which a coding assistant discovers from the skill's frontmatter and then follows step by step.[41][42] NVIDIA released them alongside JetPack 7.2, and the JetPack SDK page lists them under the heading "Jetson Agent Skills."[20][43]

The two repositories divide by where they run:

| Repository | Where it runs | Contents | Target agents |
|---|---|---|---|
| `jetson-device-skills` | On a booted Jetson | 8 skills: device diagnostics, memory audit, headless mode, inference memory tuning, LLM serving, LLM benchmarking, packaging guidance, and speculative decoding | Installer links skills into the user-level skill directories for [Claude Code](https://aiwiki.ai/wiki/claude_code), Codex, [Cursor](https://aiwiki.ai/wiki/cursor), and [NemoClaw](https://aiwiki.ai/wiki/nemoclaw) / OpenClaw sandboxes |
| `jetson-bsp-skills` | On the host workstation, before flashing | 24 skills across setup (8), customization (11), build (1), deploy (3), and inspection (1), covering pinmux, USB, PCIe, UPHY, MGBE, camera, clocks, fan, nvpmodel, and memory tuning | Described in its own README as "Claude Code skills" |

The device-side skills name the runtimes they tune for: `jetson-inference-mem-tune` recommends serving runtimes and memory flags for vLLM, SGLang, llama.cpp, and TensorRT Edge-LLM; `jetson-llm-serve` provides vLLM and SGLang serving recipes; and `jetson-llm-benchmark` emits structured metrics for vLLM, llama.cpp, and Ollama.[41] Both repositories are dual-licensed, documentation under CC BY 4.0 and source under Apache 2.0, and both state that they are not accepting contributions.[41][42] The BSP repository carries an explicit warning that flashing can erase device storage or leave a target unbootable and that generated plans, commands, and diffs should be reviewed before being accepted.[42]

### Memory as the real constraint

NVIDIA's stated reason for shipping the skills is memory rather than convenience. A coding agent can scaffold a project and wire up an inference server, NVIDIA argues, but it has never seen a particular carrier board's pinmux and cannot know which memory carveouts are safe to shrink; the skills package that board-specific knowledge into repeatable workflows.[49] NVIDIA's guidance puts total achievable savings at 10 to 12 GB across five layers of the stack, roughly 1,025 MB from board-support and OS services, about 412 MB from pipeline changes, and 5 to 10 GB from inference frameworks and model [quantization](https://aiwiki.ai/wiki/quantization), while keeping accuracy and feature parity.[49]

The commercial point of that exercise is dropping a memory tier. In its T3000 announcement NVIDIA said [UBTech](https://aiwiki.ai/wiki/ubtech), [Agile Robots](https://aiwiki.ai/wiki/agile_robots), and Connect Tech had cut memory use by up to 15 GB and moved from Jetson AGX Orin 64GB to the 32GB module; that SandStar had saved up to 4 GB in smart retail and moved from Orin NX 16GB to the 8GB part; and that NoTraffic had cut memory use by 30 percent on a Jetson TX2 NX. These are customer-reported figures relayed by NVIDIA rather than independent measurements.[43]

The Reachy Mini assistant is the worked example NVIDIA publishes for the same technique. Its Jetson AI Lab tutorial treats an Orin Nano 8GB with roughly 7.6 GB of usable unified memory as the design constraint and runs voice activity detection, speech-to-text, a vision-language model, text-to-speech, face tracking, and robot motion control inside it. The device skills diagnosed the pressure and picked the fixes: switching to a headless system target saved about 0.7 GB, and serving the Cosmos Reason2 2B vision-language model as a 4-bit GGUF quantization through llama.cpp instead of FP16 cut its footprint from about 6.6 GB to about 2.2 GB, for roughly 5 GB saved overall and a pipeline that would otherwise have needed an Orin NX 16GB-class part.[49][50]

## Why inference has to happen on the robot

NVIDIA gives four reasons for processing data where it is captured rather than in a data center: lower latency because the data travels less far, better protection for sensitive data that never leaves the device, lower bandwidth and storage cost, and reach into places with no internet access at all.[51] Each of those lands differently on a robot.

**Latency.** A humanoid's control loops run fast enough that a network round trip is not an option; NVIDIA's own humanoid guidance states that these machines "require powerful onboard compute to process multimodal data and ensure functional safety without relying on cloud connectivity," and positions Jetson AGX Thor as the on-robot runtime that supplies it, describing the module as the platform for systems "that require sophisticated real-time reasoning at the edge."[26][39]

**Bandwidth.** Sensor data volume, not model size, is often what rules out a cloud design. NVIDIA specifies the Jetson T5000 for up to 20 cameras through its Holoscan Sensor Bridge, or up to six through 16 MIPI CSI-2 lanes running D-PHY 2.1 at up to 40 Gbps, and pairs the module with a camera offload engine and 4x 25 GbE networking to move that traffic into the GPU.[44] Streams at that rate are captured to be consumed locally; what leaves the robot is the result, not the raw feed.

**Privacy and data control.** Anything processed on the module is data that never reaches a third-party server. The Reachy Mini assistant is the clearest illustration: NVIDIA describes it as running "everything locally with GPU acceleration, no cloud, no API keys, no internet required at runtime," with speech, vision, reasoning, and speech synthesis all resident on an 8GB Orin Nano.[39][50]

**Operating without connectivity.** Wireless coverage in a warehouse aisle, a field, or a disaster site is not something a robot can depend on. 1X Technologies says it selected Jetson Thor because it was the only commercially available module that met the onboard compute requirements of its home robot, which runs large models locally for perception, reasoning, and control rather than depending on a network connection.[19] The same logic drives NVIDIA's memory-optimization push: a model that fits in a module's DRAM is a model that keeps working when the link drops.[49]

## Use in humanoid and mobile robots

A humanoid robot's onboard computer has to do two very different jobs on the same board: heavy, parallel perception and planning work, interpreting camera, [lidar](https://aiwiki.ai/wiki/lidar), and [inertial measurement unit](https://aiwiki.ai/wiki/inertial_measurement_unit) data and running a model to decide what to do next, alongside fast, deterministic control loops that turn those decisions into motor commands many times per second, where a delay can mean a fall. Jetson's GPU handles the first job and its CPU cores and dedicated accelerators handle the second, on one shared board rather than two separate computers linked by a cable.[26]

NVIDIA's own May 2026 open humanoid reference design illustrates how far that integration has come: a research robot built on a Unitree H2 Plus chassis and [Sharpa](https://aiwiki.ai/wiki/sharpa) Wave tactile five-finger hands, with 75 total [degrees of freedom](https://aiwiki.ai/wiki/degrees_of_freedom) across body and hands, runs entirely from a single Jetson AGX Thor T5000 module. NVIDIA developed the design with Ai2, ETH Zurich, Stanford's Robotics Center, and UC San Diego's Advanced Robotics and Controls Laboratory for academic research, with [Unitree](https://aiwiki.ai/wiki/unitree) planning to ship it in late 2026.[27]

Adoption is not limited to Jetson's newest chip. Warehouse and delivery robots, agricultural equipment, and quadrupeds still ship largely on Orin-generation modules, which remain in production. NVIDIA's own product lifecycle listing shows Orin Nano, Orin NX, and AGX Orin available to order through January 2032, giving integrators roughly a decade of supply certainty from a single design choice; AGX Orin Industrial runs to July 2033, and the Thor modules are listed further out still, to August 2035 for T5000 and January 2036 for T4000. Most Xavier-generation modules remain available through July 2027, although the AGX Xavier 64GB variant reached end of life in January 2025, and the original Jetson Nano is scheduled to phase out in January 2027.[28]

## Alternatives to Jetson

Jetson is not the only way to put AI compute inside a robot, though independent reviewers and trade press generally describe it as the default choice for teams that are not building their own chips.[2][14] The most visible alternative is vertical integration: [Tesla](https://aiwiki.ai/wiki/tesla) designs custom silicon for its [Optimus](https://aiwiki.ai/wiki/tesla_optimus) humanoid and its vehicles rather than buying Jetson modules, on the reasoning that a chip co-designed with Tesla's own neural networks and produced at automotive volume can outperform a general-purpose part. Tesla taped out a successor chip, [AI5](https://aiwiki.ai/wiki/tesla_ai5), in April 2026; Elon Musk said at the time that it would deliver roughly 5 times the useful compute (around 8 times the raw compute) of Tesla's current in-house hardware, with volume production targeted for mid-to-late 2027 and first use in Optimus and Tesla's own compute clusters.[29] That kind of vertical integration only pays off at a scale of millions of units a year, which is part of why smaller humanoid developers have generally chosen Jetson's off-the-shelf modules and software over the cost of designing a chip from scratch.

In China, where a large share of the world's humanoid-robot manufacturing volume is now based, tightening restrictions on exports of advanced NVIDIA chips are pushing some manufacturers toward domestic accelerators such as Huawei's Ascend line. Analysts at Goldman Sachs said in a May 2026 note, reported by CNBC, that this shift toward domestic silicon in China is likely to accelerate through 2028.[30] How far that substitution extends to Jetson specifically, as opposed to the larger training chips that export controls target most directly, is not yet clear as of mid-2026.

## See also

- [Jetson Thor](https://aiwiki.ai/wiki/jetson_thor)
- [NVIDIA](https://aiwiki.ai/wiki/nvidia)
- [NVIDIA Isaac Sim](https://aiwiki.ai/wiki/nvidia_isaac_sim)
- [Isaac Lab](https://aiwiki.ai/wiki/isaac_lab)
- [Isaac GR00T](https://aiwiki.ai/wiki/isaac_gr00t)
- [Edge AI](https://aiwiki.ai/wiki/edge_ai)
- [On-device AI](https://aiwiki.ai/wiki/on_device_ai)
- [Physical AI](https://aiwiki.ai/wiki/physical_ai)
- [Humanoid robot](https://aiwiki.ai/wiki/humanoid_robot)
- [Tesla AI5](https://aiwiki.ai/wiki/tesla_ai5)
- [GPU](https://aiwiki.ai/wiki/gpu)

## References

1. NVIDIA, "Jetson AGX Orin for Next-Gen Robotics," nvidia.com, accessed July 2026. https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/
2. The Robot Report, "How does NVIDIA's Jetson Thor compare with other robot brains on the market?" 2025. https://www.therobotreport.com/how-does-nvidias-jetson-thor-compare-with-other-robot-brains/
3. NVIDIA Newsroom, "NVIDIA Blackwell-Powered Jetson Thor Now Available, Accelerating the Age of General Robotics," August 25, 2025. https://nvidianews.nvidia.com/news/nvidia-blackwell-powered-jetson-thor-now-available-accelerating-the-age-of-general-robotics
4. NVIDIA Developer, "Jetson Modules, Support, Ecosystem, and Lineup," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/embedded/jetson-modules
5. NVIDIA Technical Blog, "Introducing NVIDIA Jetson Thor, the Ultimate Platform for Physical AI," August 25, 2025. https://developer.nvidia.com/blog/introducing-nvidia-jetson-thor-the-ultimate-platform-for-physical-ai/
6. Wikipedia, "Nvidia Jetson," accessed July 2026. https://en.wikipedia.org/wiki/Nvidia_Jetson
7. NVIDIA Developer, "Jetson Download Center" and related Jetson TX1/TX2 documentation, developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/embedded/downloads
8. NVIDIA Technical Blog, "NVIDIA Jetson AGX Xavier Delivers 32 TeraOps for New Era of AI in Robotics," 2018. https://developer.nvidia.com/blog/nvidia-jetson-agx-xavier-32-teraops-ai-robotics/
9. NVIDIA Newsroom, "NVIDIA Releases Jetson Xavier NX Developer Kit with Cloud-Native Support," 2019/2020. https://nvidianews.nvidia.com/news/nvidia-releases-jetson-xavier-nx-developer-kit-with-cloud-native-support
10. NVIDIA Newsroom, "NVIDIA Announces Jetson Nano: $99 Tiny, Yet Mighty NVIDIA CUDA-X AI Computer That Runs All AI Models," March 18, 2019. https://nvidianews.nvidia.com/news/nvidia-announces-jetson-nano-99-tiny-yet-mighty-nvidia-cuda-x-ai-computer-that-runs-all-ai-models
11. NVIDIA Newsroom, "NVIDIA Announces Availability of Jetson AGX Orin Developer Kit to Advance Robotics and Edge AI," March 22, 2022. https://nvidianews.nvidia.com/news/nvidia-announces-availability-of-jetson-agx-orin-developer-kit-to-advance-robotics-and-edge-ai
12. JetsonHacks, "NVIDIA Jetson Orin Nano Developer Kit: The Perfect Solution for Makers and Developers," March 2023. https://jetsonhacks.com/2023/03/22/nvidia-jetson-orin-nano-developer-kit-the-perfect-solution-for-makers-and-developers-a-review/
13. NVIDIA Technical Blog, "NVIDIA Jetson Orin Nano Developer Kit Gets a 'Super' Boost," December 2024. https://developer.nvidia.com/blog/nvidia-jetson-orin-nano-developer-kit-gets-a-super-boost/
14. DataCenterDynamics, "Nvidia launches Jetson Thor compute modules for humanoid robots," 2025. https://www.datacenterdynamics.com/en/news/nvidia-launches-jetson-thor-compute-modules-for-humanoid-robots/
15. CNBC, "Nvidia's 'robot brain' chip, Thor, goes on sale around the world," August 25, 2025. https://www.cnbc.com/2025/08/25/nvidias-thor-t5000-robot-brain-chip.html
16. NVIDIA Developer, "Deep Learning Accelerator (DLA)," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/deep-learning-accelerator
17. The Robot Report, "NVIDIA Jetson Thor brings 2K teraflops of AI compute to robots," 2025. https://www.therobotreport.com/nvidia-jetson-thor-brings-2k-teraflops-of-ai-compute-to-robots/
18. Boston Dynamics, "Boston Dynamics Expands Collaboration with NVIDIA to Accelerate AI Capabilities in Humanoid Robots," March 18, 2025. https://bostondynamics.com/news/boston-dynamics-expands-collaboration-with-nvidia/
19. 1X Technologies, "Inside 1X's Humanoid Robot Stack: Simulation, AI Training, and Onboard Compute with NVIDIA," March 17, 2026. https://www.1x.tech/discover/nvidia-gtc-2026
20. NVIDIA Developer, "JetPack SDK," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/embedded/jetpack
21. Edge AI and Vision Alliance, "JetPack 7.2: The Production Moment for Physical AI," June 3, 2026. https://www.edge-ai-vision.com/2026/06/jetpack-7-2-the-production-moment-for-physical-ai/
22. NVIDIA Developer, "TensorRT SDK," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/tensorrt
23. NVIDIA Developer, "Isaac ROS," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/isaac/ros
24. NVIDIA, "Metropolis: Intelligent Vision AI for Smart Infrastructure," nvidia.com, accessed July 2026. https://www.nvidia.com/en-us/autonomous-machines/intelligent-video-analytics-platform/
25. NVIDIA Developer, "Isaac GR00T," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/isaac/gr00t
26. NVIDIA, "Use Case: NVIDIA Computing Platforms for Humanoid Robots," nvidia.com, accessed July 2026. https://www.nvidia.com/en-us/use-cases/humanoid-robots/
27. NVIDIA Newsroom, "NVIDIA Announces NVIDIA Isaac GR00T Reference Humanoid Robot for Academic Research," May 31, 2026. https://nvidianews.nvidia.com/news/nvidia-open-humanoid-robot-reference-design
28. NVIDIA Developer, "Jetson Product Lifecycle," developer.nvidia.com, accessed July 2026. https://developer.nvidia.com/embedded/lifecycle
29. Not a Tesla App, "Tesla's AI5 to Enter Production in 2H 2026, Rivals NVIDIA's $30K Chip in Performance," April 2026. https://www.notateslaapp.com/news/3519/teslas-ai5-to-enter-production-in-2h-2026-rivals-nvidias-30k-chip-in-performance
30. CNBC, "China learns to build without Nvidia," June 1, 2026. https://www.cnbc.com/2026/06/01/china-learns-to-build-without-nvidia.html
31. NVIDIA Developer, "JetPack 7.2," developer.nvidia.com, accessed July 25, 2026. https://developer.nvidia.com/embedded/jetpack/downloads
32. NVIDIA, "Jetson Linux 39.2.0 GA Release Notes," June 2026. https://docs.nvidia.com/jetson/archives/r39.2/ReleaseNotes/Jetson_Linux_Release_Notes_r39.2.pdf
33. NVIDIA Technical Blog, "Deploy Agentic-Ready AI at the Edge with Memory Efficiency in NVIDIA JetPack 7.2," June 1, 2026. https://developer.nvidia.com/blog/deploy-agentic-ready-ai-at-the-edge-with-memory-efficiency-in-nvidia-jetpack-7-2/
34. NVIDIA Robotics, "From a fresh install to teleoperating SO-101 robot arms in only 30 minutes," X, July 22, 2026. https://x.com/NVIDIARobotics/status/2079959655409717284
35. Back to Engineering, "NVIDIA JetPack 7.2 Full Setup: PyTorch, ROS 2 & LeRobot in 30 Minutes (2026)," YouTube, July 15, 2026. https://www.youtube.com/watch?v=1Qb_foDfCwo
36. Iulia Feroli, "PyTorch on JetPack 7.2 for Jetson Orin NX (the undocumented path)," GitHub, accessed July 25, 2026. https://github.com/iuliaferoli/jetson-jp7.2-install
37. NVIDIA, "PyTorch for Jetson Platform: Compatibility," updated July 9, 2026. https://docs.nvidia.com/deeplearning/frameworks/install-pytorch-jetson-platform-release-notes/pytorch-jetson-rel.html
38. Hugging Face, "LeRobot SO-101," accessed July 25, 2026. https://huggingface.co/docs/lerobot/en/so101
39. NVIDIA Blog, Matthew Leib, "Powerful Compute So Compact, It's Clutch: Build AI Anywhere With NVIDIA Jetson," July 28, 2026. https://blogs.nvidia.com/blog/build-ai-with-nvidia-jetson/
40. NVIDIA Robotics, "Bring classic games to life with GenAI on NVIDIA Jetson," X, July 27, 2026. https://x.com/NVIDIARobotics/status/2081844348128243952
41. NVIDIA-AI-IOT, "jetson-device-skills," GitHub, accessed July 31, 2026. https://github.com/NVIDIA-AI-IOT/jetson-device-skills
42. NVIDIA-AI-IOT, "jetson-bsp-skills: Agentic Skills for Jetson BSP Customization," GitHub, accessed July 31, 2026. https://github.com/NVIDIA-AI-IOT/jetson-bsp-skills
43. NVIDIA Blog, Chen Su, "NVIDIA Introduces New Jetson Thor Computers to Advance Mainstream Robotics and Edge AI," July 15, 2026. https://blogs.nvidia.com/blog/jetson-thor-robotics-edge-ai-agent/
44. NVIDIA, "NVIDIA Jetson Thor" product page and module specifications, nvidia.com, accessed July 31, 2026. https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-thor/
45. NVIDIA Technical Blog, Shashank Maheshwari and Chen Su, "NVIDIA JetPack 6.2 Brings Super Mode to NVIDIA Jetson Orin Nano and Jetson Orin NX Modules," January 16, 2025. https://developer.nvidia.com/blog/nvidia-jetpack-6-2-brings-super-mode-to-nvidia-jetson-orin-nano-and-jetson-orin-nx-modules/
46. NVIDIA, "Jetson AGX Orin Series Technical Brief," TB_10749-001_v1.2, July 2022. https://www.nvidia.com/content/dam/en-zz/Solutions/gtcf21/jetson-orin/nvidia-jetson-agx-orin-technical-brief.pdf
47. NVIDIA, "Jetson Orin NX Series Data Sheet," DS-10712-001, accessed July 31, 2026. https://developer.nvidia.com/downloads/jetson-orin-nx-series-data-sheet
48. NVIDIA, "Jetson Orin Nano Super Developer Kit," nvidia.com, accessed July 31, 2026. https://www.nvidia.com/en-us/autonomous-machines/embedded-systems/jetson-orin/nano-super-developer-kit/
49. NVIDIA Technical Blog, Anshuman Bhat and Aditya Sahu, "Maximizing Memory Efficiency with Agent Skills to Run Bigger Models on NVIDIA Jetson," accessed July 31, 2026. https://developer.nvidia.com/blog/maximizing-memory-efficiency-to-run-bigger-models-on-nvidia-jetson/
50. NVIDIA Jetson AI Lab, Aditya Sahu, "Reachy Mini Jetson Assistant," accessed July 31, 2026. https://www.jetson-ai-lab.com/tutorials/reachy-mini-jetson-assistant/
51. NVIDIA, "Edge Computing," nvidia.com, accessed July 31, 2026. https://www.nvidia.com/en-us/edge-computing/