At COMPUTEX, NVIDIA unveiled JetPack 7.2 and expanded NemoClaw support for its Jetson platform, marking a significant shift toward deploying agentic AI in the physical world. This release transitions advanced AI capabilities from servers to edge devices, targeting robotics, industrial automation, and inspection systems. Deepu Talla, NVIDIA vice president of robotics and edge computing, emphasized that Jetson's programmability and performance now enable developers to deploy production-ready physical AI agents immediately. JetPack 7.2 introduces three critical layers to the ecosystem. The foundation features a Yocto-based operating system, offering a lean, customizable Linux environment ideal for memory-constrained industrial deployments. It also includes CUDA 13 for the Jetson Orin line and Multi-Instance GPU support on the Jetson Thor, allowing developers to reserve dedicated GPU resources for deterministic workloads like robot perception. Additionally, the Jetson AGX Orin 32GB module sees a 20% performance increase, reaching 241 TOPS of AI compute. The middle layer introduces agent skills that automate developer tasks such as Linux customization, memory optimization, and model benchmarking. These tools, derived from NVIDIA documentation, aim to reduce development time from weeks to days. At the apex is NemoClaw, an agentic AI framework that can be deployed to Jetson with a single command. This integration allows developers to rapidly build systems with visual reasoning agents capable of interpreting and acting on visual data, supported by NVIDIA Metropolis VSS blueprint skills. Real-world applications are already emerging across various sectors. Solomon leverages NemoClaw to coordinate AI agents in humanoid robots, integrating perception, locomotion, and manipulation into unified workflows for complex environments. Advantech is implementing an agentic factory brain using Jetson Thor to automate robot fleet management and defect detection. In the smart city sector, Rebotnix has developed cameras with agentic reasoning to accelerate urban decision-making, while Spingence uses the technology to identify manufacturing root causes. Efficiency gains are a major driver for these deployments. SandStar reduced memory requirements by nearly 40% for its AI vending machines, allowing migration from 16GB to 8GB hardware without sacrificing performance. NoTraffic optimized its traffic management systems by pruning CUDA kernels, achieving a 29% reduction in memory usage to improve real-time inference. GROOVE X similarly reduced its memory footprint for its LOVOT companion robot by offloading workloads to AI accelerators. Yocto-based customization is proving essential for safety and reliability in demanding environments. Hexagon Robotics utilizes Jetson Thor with Yocto for safer humanoid robots in manufacturing and logistics. Zipline employs Yocto for its autonomous delivery drones to ensure high-performance sensor fusion and safe navigation. Major hardware partners including 1X, Universal Robots, Balena, and Wind River are validating these Yocto-based deployments to accelerate market entry. This launch signifies the beginning of an era where physical AI agents operate reliably at the edge. Developers are now invited to access the Jetson software page and NVIDIA's GTC Taipei recordings to explore these new capabilities further.