At this year's re:Invent 2025 conference, in addition to the arms race in cloud computing power and generative AI models, AWS has also set its sights on the field of "Physical AI," which possesses perception and interaction capabilities. Through in-depth cooperation with NVIDIA and the continued promotion of the "Physical AI Fellowship" competition program, AWS is attempting to extend its cloud advantages to real-world robotics applications.
In the "Physical AI Showcase" area, AWS showcased six representative startups covering vertical fields such as healthcare, aerospace, industrial automation, agriculture, and marine exploration, demonstrating how AWS's technology stack empowers the development of physical AI.
Diligent Robotics: Social Robots in Hospitals
Diligent Robotics has created AI robots with social intelligence, aiming to automate hospital workflows and reduce the burden on nursing teams.
AWS technology applications:The basic model for developing autonomous humanoid robots is developed using AWS computing power and AI stacks, including its first VLA (visual-language-action) model for "pressing elevator buttons".
Fellowship Program Outcomes:Through the Physical AI Fellowship competition program, the team optimized the architecture, solved data filtering artifacts, and successfully deployed Amazon SageMaker HyperPod orchestrated by Slurm to achieve scalable experimental training.
Voxelis AI: Helicopters Transform into AI Firefighting Vanguard
Founded by aerospace and technology veterans, Voxelis AI is dedicated to transforming traditional helicopters into autonomous "firefighting helicopters" with pilot-level capabilities.
AWS technology applications:Its VoxVision platform combines AWS cloud, IoT and edge AI technologies to support real-time thermal imaging scanning and environmental monitoring, and can operate smoothly even in wildfire sites with low visibility and strong winds.
Data integration:Leveraging AWS processing power, it automatically organizes data with precise geotags and timestamps, enabling seamless collaboration between aircrew and fire departments.
RobCo: A Driving Force for the Popularization of Modular Robots
RobCo combines patented modular hardware with a "no-code" workshop to make industrial robots affordable for small and medium-sized enterprises.
AWS technology applications:To scale from single-node to multi-node training, RobCo deployed Amazon SageMaker HyperPods on Amazon EKS using AWS, along with Karpenter autoscaling.
Fellowship Outcomes:Create Ray clusters using the KubeRay operator to enable efficient distributed training and rapid version iteration updates on AWS.
Roomie IT: Proactive Computer Vision Protection
Roomie IT’s ROI First Enterprise AI platform includes eight agent modules, among which the Physical AI Agent transforms computer vision from passive detection to proactive prevention.
AWS technology applications:Real-time capture is performed using AWS Kinesis Video Streams, the visual model is trained/deployed using Amazon SageMaker AI, and the multimodal underlying model is accessed via Amazon Bedrock. Finally, secure and scalable edge execution is achieved using IoT Greengrass.
Aigen: Solar-Powered Autonomous Weeding Robot
Aigen's solar-powered, fully autonomous weeding robot aims to eliminate the need for herbicides and address labor shortages on farms.
AWS technology applications:Through AWS Compute for Climate Fellowship, Aigen expands its ability to collect and analyze massive datasets, accelerating the training of basic models and enabling its robots to adapt to different crops.
Operation mode:The robot team communicates through a smart mesh system, enabling it to operate 24/7 and use AI vision to precisely remove weeds.
COSMA: Digital Twin of Deep-Sea Data
COSMA uses autonomous underwater drone swarms to capture high-resolution images, providing deep-sea data for environmental impact assessments and infrastructure inspections.
AWS technology applications:By leveraging an AI platform to generate 3D seabed reconstructions and species identification, a complete deep-sea digital twin is constructed. As a member of the AWS Compute for Climate Fellowship, COSMA utilizes Amazon EKS's automatic scaling of node groups to easily handle sudden surges in computational loads.
Analysis: AWS's deployment of physical AI hinges on the key integration of hardware and software.
These cases demonstrate that AWS's strategy for physical AI is not to build robots itself, but rather to play the role of "the most powerful brain and nervous system".
• Computing power and model support:Through SageMaker HyperPod and Amazon Bedrock, AWS solves the most expensive model training and inference challenges for robotics startups.
• Edge and cloud collaboration:By leveraging IoT Greengrass and Kinesis, a closed loop is achieved, from edge data collection and cloud training to edge execution.
• Supported by NVIDIA collaboration:Combined with NVIDIA's Isaac robotics simulation and development platform, AWS provides a complete toolchain from virtual simulation to real deployment, which is the core value of the Physical AI Fellowship competition program.
As AI moves from the digital world to physical interaction, AWS is using these startup partners to validate the ability of its cloud infrastructure to be deployed in the real physical world.
