From Software Agents to Physical AI on NVIDIA Jetson
Agentic AI deployment on NVIDIA Jetson means software agents that plan, decide, and act can run on compact edge computers to control robots, inspection systems, and other physical machines in real time. Instead of living only in cloud data centers, these AI agents move into the physical world, where they handle perception, planning, and actuation close to sensors and motors. NVIDIA JetPack 7.2 is the latest software foundation enabling this shift, pairing the NVIDIA Jetson edge AI hardware family with a memory-efficient, production-grade stack for robotics and industrial automation. By combining an updated compute stack, deterministic performance features, and purpose-built agent skills, JetPack 7.2 turns existing Jetson Orin and upcoming Jetson Thor modules into platforms where AI agents can safely run complex workflows at the edge, with less dependence on remote servers.
JetPack 7.2: Agentic-Ready, Memory-Efficient Edge AI
The JetPack 7.2 release is designed to make NVIDIA Jetson edge AI platforms agentic-ready while improving memory efficiency for constrained deployments. Out of the box, JetPack 7.2 ships with NVIDIA CUDA 13 on Jetson Orin and a refreshed OS and compute stack, giving existing devices a new performance baseline. It also introduces Yocto Project support so teams can build lean, custom Linux images, an important gain for memory-bound edge computing robotics and inspection systems. According to NVIDIA, Jetson AGX Orin 32GB now reaches 241 TOPS of AI compute, a 20% uplift over its original specification, while a new Super Mode improves cost efficiency at the edge. These updates help developers run more demanding AI agents physical world workloads on lower-memory configurations, extending the usable life of deployed hardware and reducing total cost of ownership across fleets.
NemoClaw and Agent Skills: A Production Stack for AI Agents
JetPack 7.2 brings NVIDIA NemoClaw directly onto the Jetson production stack, aligning the framework with real-world intelligence deployment. NemoClaw, an open-source agentic AI framework, now installs on Jetson with a single command and arrives pre-integrated with required dependencies. This allows developers to stand up agentic AI deployment pipelines for robotics, industrial automation, and vision agents without fragile custom environments. In parallel, NVIDIA Jetson agent skills formalize repeatable workflows as agent-executable instructions, spanning Jetson Linux customization, memory optimization, and model benchmarking. Tasks like configuring custom BSPs, tuning bootloader memory carveouts, or benchmarking inference now become automated jobs that AI agents can run. The result is a three-layer stack: JetPack 7.2 at the base, agent skills in the middle, and NemoClaw on top, turning Jetson into a practical platform for deploying, tuning, and operating AI agents in the physical world.
Deterministic Edge Computing for Robotics and Industrial Automation
To support time-sensitive edge computing robotics scenarios, JetPack 7.2 adds features aimed at predictable, mixed-criticality workloads. On NVIDIA Jetson Thor, Multi-Instance GPU (MIG) support can split the integrated NVIDIA Blackwell GPU into two isolated instances, each with its own compute, cache, and memory bandwidth. Combined with a preemptible real-time kernel, this allows developers to reserve deterministic GPU capacity for critical tasks like robot perception or industrial inspection, while less urgent AI agents share remaining resources. At the system level, official Yocto support and memory optimization skills help strip away unnecessary user-space processes and tune kernel reservations, reducing latency and jitter. This edge-first design reduces dependency on cloud infrastructure, keeping perception, decision, and actuation loops local so AI agents can respond in real time, even in environments with unreliable or bandwidth-limited connectivity.