AWS re:Invent 2025 - Train, Simulate, Deploy: Building NVIDIA-Powered Physical AI on AWS (AIM117)

Building NVIDIA-Powered Physical AI on AWS

Overview of Physical AI

• Physical AI is the embodiment of AI in physical objects and smart devices
• Key industries that can benefit include manufacturing, supply chain, logistics, healthcare, agriculture, retail, and public sector
• NVIDIA CEO estimates 65% of global GDP can be improved by physical AI, a $50+ trillion economic impact

Key Pillars of Physical AI

1. Data: Requires massive amounts of high-quality, realistic data - much of which is generated synthetically
2. Training: Leverages techniques like fine-tuning, imitation learning, and reinforcement learning
3. Simulation: Tests trained models extensively in a simulated environment before real-world deployment
4. Sim-to-Real: Bridges the gap between simulation and real-world performance through iterative feedback
5. Agentic Orchestration: Combines modular AI agents to enable complex, end-to-end behaviors

Challenges and Solutions

• High data requirements - Addressed by NVIDIA Cosmos models and synthetic data generation
• Simulation accuracy - NVIDIA Omniverse provides physically accurate, photorealistic simulation
• Complexity of tools - NVIDIA NIMS packages solutions for easy deployment and repeatability
• Collaboration and standards - Enables sharing and reuse across teams and locations
• Training time and cost - Leverages multi-GPU/multi-node parallelization and cloud elasticity

Example: Humanoid Stockkeeping Robot

• Developed using the 5 pillars and NVIDIA/AWS technologies
• Captures 800 episodes of human teleoperation data as baseline
• Augments with synthetic data using NVIDIA Cosmos models
• Trains policies using NVIDIA Isaac Lab on AWS Batch
• Validates in simulation using NVIDIA Isaac Sim on AWS
• Deploys to edge using NVIDIA Jetson devices and manages fleet with AWS IoT
• Integrates modular AI agents built with Amazon Bedrock for complex behaviors

Key Takeaways

• Physical AI is a transformative technology with massive business potential across industries
• NVIDIA and AWS provide a comprehensive, integrated ecosystem to accelerate physical AI development
• Synthetic data generation, photorealistic simulation, and modular AI agents are critical enablers
• Iterative, cloud-based workflows can dramatically improve time-to-results and cost-effectiveness
• Real-world examples like the humanoid stockkeeping robot demonstrate the power of this approach