The Cat 306 CR mini-excavator weighs slightly below eight tons and suits inside a normal delivery container. It’s the machine a contractor rents when the job web site is tight: a utility trench close to a basis, a basement dig in a dense neighborhood.
The cab is roughly the scale of a telephone sales space. The operator sits near the controls, two joysticks, a number of features per hand. It takes time to be taught. It takes longer to hurry up.
At CES earlier this yr, that machine answered questions.
Within the demo, the Cat AI Assistant ran on NVIDIA Jetson Thor, an edge AI platform constructed for actual‑time inference in industrial and robotic programs, NVIDIA Nemotron speech fashions are used for quick and correct pure voice interactions. Qwen3 4B, served regionally by way of vLLM, interprets requests and generates responses with low latency, no cloud hyperlink required.
Past enterprise innovation, open fashions unlock new prospects for builders to construct and experiment freely. Operating OpenClaw on NVIDIA Jetson allows builders to create non-public, always-on AI assistants on the edge — with zero utility programming interface price and full information privateness.
All Jetson developer kits assist OpenClaw, providing the pliability to change throughout open fashions from 2 billion parameters to 30 billion. With a frontier-class AI assistant operating regionally, customers can energy morning briefings, automate each day duties, carry out code critiques and management good dwelling programs — all in actual time.
From the Cloud to the Edge
For many of their current historical past, open fashions lived the place it was best to assist them.
They ran in information facilities, backed by elastic compute and chronic networks. Cloud deployments carry prices in latency and ongoing compute spend that scale with each question.
Bodily programs optimize for one thing else. Low latency as a result of machines work together with individuals and environments. Restricted energy as a result of units have arduous limits. And constant habits as a result of variability introduces danger.
There’s additionally a provide query. Reminiscence shortages have pushed up prices throughout the {industry}. Jetson brings compute and reminiscence collectively in a system-on-module, accelerating buyer {hardware} design and making sourcing and validation simpler than with discrete part approaches.
And as fashions have grown extra environment friendly, builders have additionally began asking a distinct query. Not which mannequin performs greatest in isolation, however the place it is smart to run.
Extra typically, the reply is on the gadget, ranging from Jetson Orin Nano 8GB for entry-level generative AI fashions.
Constructing Autonomous Bodily AI Programs at Scale
For bodily AI programs, generative AI fashions are increasing what’s attainable.
Caterpillar’s in-cab Cat AI Assistant, which is in growth, runs speech and language fashions regionally alongside trusted machine context, supporting operator steerage and security options.
At CES, Franka Robotics confirmed what that appears like in robotics. The firm’s FR3 Duo dual-arm system ran the NVIDIA GR00T N1.6 mannequin end-to-end onboard, notion to movement, no activity scripting. The coverage executes regionally.
In robotics analysis, the SONIC challenge from NVIDIA’s GEAR Lab trains a humanoid controller on over 100 million frames of motion-capture information, then deploys the ensuing coverage on a bodily robotic the place the kinematic planner runs on Jetson Orin at round 12 milliseconds per go. The coverage loop runs at 50 Hz. Every thing executes onboard.
The sample reaches into the developer neighborhood. A group from UIUC’s SIGRobotics membership constructed a dual-arm matcha-making robotic on Jetson Thor operating the GR00T N1.5 mannequin. It took first place at an NVIDIA embodied AI hackathon.
This analysis momentum continues on the NYU Heart for Robotics and Embodied Intelligence. The group not too long ago ran its YOR robotic on Jetson Thor, utilizing NVIDIA Blackwell compute to deal with the heavy processing required for AI-driven motion. Early outcomes present YOR performing intricate pick-and-place duties with higher generalization to new objects and robustness to scene variation, accelerating readiness for a variety of family duties like cooking and laundry.
Impartial researchers are discovering the identical. Andrés Marafioti, a multimodal analysis lead at Hugging Face, constructed an agentic AI system on Jetson AGX Orin that routes duties throughout fashions and schedules its personal work. Late one night time, the agent despatched him a message: Fall asleep. Every thing will likely be prepared by morning.
Developer Ajeet Singh Raina from the Collabnix neighborhood has proven the way to run OpenClaw on NVIDIA Jetson Thor for a private AI assistant that runs 24/7. This setup permits for personal massive language mannequin inference for the consumer’s personal information whereas the system manages emails and calendars by a neighborhood gateway.
Jetson Is the New Normal
NVIDIA Jetson has change into a typical platform for operating open fashions on the edge.
It helps a variety of open fashions and AI frameworks, giving builders flexibility for nearly any generative AI workload on the edge.
Mannequin benchmarks can be found at Jetson AI Lab, together with tutorials from the open mannequin neighborhood. Jetson Thor delivers management inference efficiency throughout all main generative AI fashions.
Gemma: Constructed on Google’s Gemini analysis, Gemma 3 is a flexible workhorse for Jetson. It’s multimodal out of the field, which suggests it may possibly see and speak in over 140 languages. On Jetson Thor, it handles a large 128K context window. This makes it good for robots that want to recollect a protracted record of advanced or multistep directions.
gpt-oss-20B: This mannequin from OpenAI lowers the barrier to deploying superior AI by delivering close to state-of-the-art reasoning efficiency in a mannequin that may run regionally on Jetson Thor and Orin for cost-efficient inference.
Mistral AI: The brand new Mistral 3 open mannequin household delivers industry-leading accuracy, effectivity and customization capabilities for builders and enterprises. This household consists of small, dense fashions starting from 3B to 14B, quick and remarkably good for his or her dimension. Jetson builders can use the vLLM container on NVIDIA Jetson Thor to attain 52 tokens per second for single concurrency, with scaling as much as 273 tokens per second with concurrency of eight.
NVIDIA Cosmos: This main, open, reasoning imaginative and prescient language mannequin allows robots and AI brokers to see, perceive and act within the bodily world like people. Each the 8B and 2B fashions run on Jetson to ship superior spatial-temporal notion and reasoning capabilities.
NVIDIA Isaac GR00T N1.6 is an open imaginative and prescient language motion mannequin (VLA) for generalist robotic expertise. Builders can use it to construct robots that understand their atmosphere, purpose about directions and act throughout a variety of duties, environments and embodiments. On Jetson Thor, the complete GR00T N1.6 pipeline executes onboard, delivering real-time notion, spatial consciousness and responsive motion.
NVIDIA Nemotron: A household of open fashions, datasets and applied sciences that empower customers to construct environment friendly, correct and specialised agentic AI programs. It’s designed for superior reasoning, coding, visible understanding, agentic duties, security, speech and knowledge. The Nemotron 3 Nano 9B mannequin successfully runs on Jetson Orin Nano Tremendous with llama.cpp with 9 tokens per second efficiency.
PI 0.5: A VLA mannequin from Bodily Intelligence that allows robots to grasp directions and autonomously execute advanced real-world duties with sturdy generalization and real-time adaptability, whereas NVIDIA Jetson Thor delivers 120 motion tokens per second to energy responsive, low-latency bodily AI deployment.
Qwen 3.5: This household of fashions from Alibaba, together with the most recent Qwen 3.5 releases, provides a mixture of dense and combination‑of‑consultants fashions that ship sturdy reasoning, coding multimodal understanding and lengthy‑context efficiency. Jetson Thor delivers optimized efficiency throughout Qwen fashions just like the Qwen 3.5-35B-A3B mannequin, which causes at 35 tokens per second, making real-time interactivity attainable.
Any developer can fine-tune these fashions to create specialised bodily AI brokers and seamlessly deploy them into bodily AI programs. The NVIDIA Jetson platform helps widespread AI frameworks from NVIDIA TRT, Llama.cpp, Ollama, vLLM, SGLang and extra.
Take On Open Fashions on Jetson
Builders can dive into Hugging Face tutorials — together with Deploying Open Supply Imaginative and prescient Language Fashions on Jetson — and catch the most recent livestream. Study from this tutorial and run OpenClaw on NVIDIA Jetson.
Be a part of GTC 2026 subsequent month to see all of it in motion. NVIDIA will present how open fashions are shifting from information facilities into machines working within the bodily world, together with in a panel on the Way forward for Industrial Autonomy.
Watch the GTC keynote from NVIDIA founder and CEO Jensen Huang and discover bodily AI, robotics and imaginative and prescient AI classes.
