Building intelligent physical AI: From edge to cloud with Strands Agents, Bedrock AgentCore, Claude 4.5, NVIDIA GR00T, and Hugging Face LeRobot

This article explores building intelligent physical AI systems that combine edge computing for real-time control with cloud computing for complex reasoning, using Strands Agents, Bedrock AgentCore, and robotics platforms.

• Physical AI requires millisecond-level edge responses for sensing/actuation and cloud reasoning for planning
• Strands Agents SDK now supports TypeScript, evaluations, bidirectional streaming, and steering capabilities
• Demonstrations show SO-101 robotic arm with NVIDIA GR00T and Boston Dynamics Spot controlled via unified interface
• Edge agents run open-source models (Qwen3-VL) locally using Ollama for low-latency vision understanding
• Vision-Language-Action models like GR00T enable robots to perceive, reason, and act in physical environments
• Agents-as-tools pattern allows edge devices to delegate complex reasoning to cloud-based agents
• AgentCore Memory enables fleet-wide learning and collective intelligence across multiple robots
• AgentCore Observability provides tracing and insights for continuous improvement across robot fleets
• SageMaker enables parallel simulation and training to apply learnings back to improved models

The convergence of powerful multimodal models, edge hardware, and open-source robotics makes physical AI development increasingly accessible for building autonomous systems that learn and improve at scale.