From Software Agents to Physical AI Systems
NVIDIA’s new AI agent stack for robotics and industry is an integrated platform that combines open-source physical AI tools, edge-ready software, and long-running reasoning models to let autonomous agents plan, decide, and act in real-world environments over many steps while staying efficient and secure. Announced around GTC Taipei, the stack moves AI agents beyond code generation into areas like robotics, autonomous vehicles, and industrial automation agents. NVIDIA has open-sourced a collection of skills and tools across Omniverse, Cosmos, Isaac, Metropolis, Alpamayo, and Jetson so AI agents can execute repeatable workflows for simulation, synthetic data generation, training, validation, and deployment. Jensen Huang said that when agents can directly use NVIDIA libraries, models, and frameworks, development of physical AI systems for transportation, manufacturing, healthcare, and robotics accelerates, cutting both cost and complexity for AI agents robotics projects.
JetPack 7.2 Brings Agentic AI to the Edge
NVIDIA JetPack 7.2 is the software foundation that brings NVIDIA JetPack agentic AI to edge devices, turning Jetson-based robots and industrial systems into capable, long-running AI agents. The release adds agentic AI skills, Yocto Project support, and NVIDIA CUDA 13 on Jetson Orin, plus Multi-Instance GPU on Jetson Thor, aligning edge AI deployment with data center practices. Yocto support gives industrial teams a lean, customizable Linux base, important for memory-bound agent workloads that must run for hours or days in constrained devices. According to NVIDIA, “Jetson’s programmability and high performance enable developers to instantly deploy physical AI agents in production at the edge.” Three layers define the stack: JetPack 7.2 at the base for deterministic performance, a middle layer of agent skills to automate developer tasks, and the NemoClaw framework on top for orchestration across robotics and inspection systems.
NemoClaw and Nemotron 3 Ultra for Long-Running Agents
At the orchestration layer, NVIDIA NemoClaw defines an open blueprint for building Nemotron long-running agents that coordinate tools, sub-agents, and complex workflows. NemoClaw works with multiple orchestration harnesses such as OpenClaw and Hermes, and uses NVIDIA NeMo libraries and a model router so enterprises can mix different models for different tasks. Nemotron 3 Ultra sits at the reasoning core: a 550B-parameter Mixture-of-Experts model with 55B active parameters designed for frontier planning. It maintains context across many turns, supports million-token contexts, and achieves up to 5x higher throughput than comparable open models, which lowers agent cost by up to 30% through fewer tokens per task. This combination lets AI agents robotics applications sustain long engineering sessions, coordinate tool chains, and avoid goal drift while still meeting performance and cost targets for production workloads.
Industrial Automation Agents Compress Weeks of Work into Hours
Industrial software vendors are treating NVIDIA’s stack as a way to build autonomous AI engineers that automate end-to-end engineering workflows. Accelerated computing already cuts core simulation times from weeks to hours, but setup and post-processing—CAD, meshing, case configuration, debugging, and reporting—remain labor-intensive. With NemoClaw and Nemotron long-running agents, companies in automotive, aerospace, semiconductors, and manufacturing are building industrial automation agents that span these steps. Cadence, for example, is developing an autonomous RTL engineer that orchestrates its ChipStack design and verification flow, reducing RTL verification time from weeks to hours. Enterprises can deploy these agents from DGX Spark personal AI supercomputers, in data centers, or through cloud providers, and then connect them to Jetson-powered equipment on the factory floor, turning simulation decisions into physical actions in robotics cells and inspection lines.
Security, Cost and Scale for Enterprise Edge AI Deployment
For enterprises, the appeal of this stack is less about demos and more about secure, scalable edge AI deployment. NemoClaw’s core runtime, NVIDIA OpenShell, defines how each agent accesses files, networks, and external tools, enforcing policy-based security at every layer. That is vital when AI agents control physical robots, inspection systems, or industrial digital twins connected to production networks. JetPack 7.2’s memory-optimized platform and Yocto-based OS help teams ship reliable, always-on agents on Jetson Orin and future Jetson Thor systems, keeping hardware costs and power budgets in check. Nemotron 3 Ultra reduces token usage and increases throughput, which lowers operational costs as agents scale across lines and sites. Together, these pieces form a unified platform so enterprises can roll out AI agents robotics workloads—from design offices to factory floors—without rebuilding their stack for each environment.
