From Software Agents to Physical AI on Jetson
Agentic AI on NVIDIA Jetson is the use of autonomous, tool-using AI agents that plan, act and adapt on-device to control real robots and industrial systems in the physical world, instead of remaining limited to purely digital or cloud-based tasks. With the latest NVIDIA Jetson edge AI platform, this idea is moving from concept to deployment. The JetPack 7.2 release introduces a software stack that lets AI agents run close to sensors and actuators, where low latency and reliability matter most. Three layers define this stack: JetPack 7.2 as the base operating system and compute layer, a set of agent skills that automate development workflows, and NemoClaw at the top as a production agentic framework. Together, they aim to make agentic AI deployment practical in robotics, inspection, industrial automation and vision-driven edge applications.
JetPack 7.2: Agentic-Ready and Memory-Efficient at the Edge
The JetPack 7.2 release focuses on making agentic AI deployment efficient on memory-constrained edge devices. It brings CUDA 13 to NVIDIA Jetson Orin and a performance uplift on Jetson AGX Orin 32 GB, which now reaches 241 TOPS of AI compute, while adding Super Mode for more performance and better cost efficiency at the edge. JetPack 7.2 also introduces official Yocto Project support, letting industrial teams build lean, custom Linux distributions that reduce overhead and improve determinism. On NVIDIA Jetson Thor, Multi-Instance GPU (MIG) support and a real-time kernel enable deterministic multiworkload execution, so critical perception or control tasks can have reserved GPU resources. According to NVIDIA, this multi-generation, software-defined platform means existing Jetson hardware gains new capabilities with each release, helping developers shorten time to market and lower total cost of ownership for edge robotics AI deployments.
Open-Source Physical AI Tools Accelerate Robotics and Industry
Alongside JetPack 7.2, NVIDIA has open sourced a broad collection of physical AI agent tools and skills spanning Omniverse, Cosmos, Isaac, Metropolis, Alpamayo and Jetson technologies. These tools turn complex steps such as simulation, synthetic data generation, training, validation and deployment into repeatable workflows that AI agents can execute. For robotics and industrial automation teams, this reduces manual integration work and encourages simulation-driven design and AI-enabled manufacturing workflows. By giving agents direct access to NVIDIA libraries, models and frameworks, developers can build and scale robots, autonomous vehicles, digital twins and vision AI systems more quickly. The goal is to cut the cost and complexity of edge robotics AI development while making AI agents capable of handling end-to-end tasks, from virtual testing to real-world deployment on the factory floor or in inspection scenarios.
NemoClaw: Production Stack for Agentic AI Deployment on Jetson
NemoClaw is NVIDIA’s agentic AI framework, and JetPack 7.2 makes Jetson NemoClaw-ready out of the box. The platform comes preconfigured with all required dependencies so developers can deploy NemoClaw-based workflows on Jetson with a single command, without manual environment setup. NemoClaw adds privacy and security controls to the open source OpenClaw stack, providing a production-grade path for deploying agentic AI in robotics, inspection and industrial automation. On-device NemoClaw agents can coordinate perception, planning and control while interacting with NVIDIA software libraries. This stack allows AI agents to operate independently at the edge, near cameras, sensors and actuators, which cuts latency and reduces reliance on data-center connectivity. That independence is vital for use cases such as inspection lines, mobile robots or medical devices that must keep running even when network links are unreliable or unavailable.
Agent Skills and Edge Independence for Industrial AI
JetPack 7.2 ships a new layer of NVIDIA agent skills for Jetson, which are reusable, agent-executable instructions that describe how to perform development and optimization tasks. Three categories cover key workflows: Jetson Linux customization skills for configuring BSPs, I/O and power profiles on custom carrier boards; memory optimization skills for tuning bootloader carveouts, kernel reservations and user-space processes; and model benchmarking skills for selecting and optimizing models for specific Jetson targets. These skills let AI agents automate steps that previously took weeks of manual engineering. Deployed at the edge, agentic AI can use these skills to adapt systems over their lifetime, improving performance or efficiency without extensive human intervention. Combined with the open-source physical AI tools and the NemoClaw stack, Jetson enables practical, low-latency agentic AI deployment in inspection, manufacturing and broader industrial AI environments.