How Agility Robotics scales AI model training for next-generation humanoid robots using AWS

Agility Robotics Robotics needed a solution to scale their AI model training capabilities beyond the constraints of on-premises infrastructure.  According to Johnson, "We were limited by GPU availability and infrastructure reliability issues on-premises. To advance our robotics capabilities, we needed consistent access to specialized GPU resources with specific requirements around clock speed and memory."

Traditional on-premises infrastructure could not provide the flexibility and scale needed for training complex perception and reinforcement learning models. The company required a cloud solution that could support their unique GPU requirements for robotics workloads.

To accelerate development of Digit's capabilities, Agility Robotics chose to migrate their training workloads to AWS, leveraging specialized GPU instances and comprehensive cloud services.  They knew that cost-effective model training would be critical to the evolution of Digit’s capabilities, and needed a solution that could focus on both reinforcement learning and perception learning, "We train Digit in simulation using tens of thousands of parallel scenarios. AWS provides the infrastructure flexibility and scale we need to advance our robots' abilities to walk, run, and navigate uneven terrain," said Johnson.  

Humanoids are at the forefront of applied AI, with development timelines increasingly compressed. To keep pace, Agility implemented a comprehensive AWS solution that integrates multiple services to support their robotics development pipeline. The company utilizes AWS EC2 G6 GPU instances for intensive model training and simulation workloads. When deployed at a customer site such as a warehouse, Digit can run into a variety of unexpected situations, just as a person would, adapting to changing environments. By building its models on AWS, Agility can always be improving Digit’s ability to react to new situations based on new data.

Agility’s architecture builds in a variety of AWS solutions, leveraging Amazon Simple Storage Service (Amazon S3) for efficient data storage and using the Amazon Relational Database Service (RDS) for full database management. To build, scale, and operate the production-grade generative AI applications necessary to power Digit, the company relies on Amazon Elastic Kubernetes Service (EKS), which allows Agility to customize their workloads, implement the tenant isolation their applications need, and orchestrate temporary jobs easily, all while managing training and inference deployments and reducing infrastructure complexity and costs. "AWS allows us to improve Digit's perception capabilities through massive-scale simulations, resulting in better walking, running, and terrain navigation," said Johnson.

AWS has transformed Agility Robotics' development capabilities, expanding the scope of AI from artificial intelligence to artificial dexterity by empowering Digit’s ever-expanding capabilities. They accomplished this by migrating their previously fragmented service infrastructure from on-premises to AWS. Through scalable GPU access with Amazon EC2, Agility has dramatically increased its model training capacity, leading to an ever-accelerating evolution of robot perception and movement capabilities. The massive amounts of trainable data needed for Digit’s optimal performance relies on Amazon S3’s scalability, data availability, durability, and security.  And the scalability of the cloud has helped Agility drive cost and energy efficiencies through optimizing resource use, avoiding over-provisioning, and freeing teams to focus on making Digit more capable and versatile.

Agility’s AWS infrastructure has supported the implementation of enhanced safety features such as controlled orderly stops, often through electronic braking. and safety Programmable Logic Controller (PLC) integration, which can detect hazards and ensure fail-safe operation. Additionally, AWS has worked with Agility to explore partnership opportunities with major customers, such as companies in the automotive industry. The company continues to expand what Digit can achieve while maintaining its focus on worker safety and the growing potential for productive collaboration between humans and robots.

“In a field cluttered with ‘demo-ware’ and hype, Agility stands apart for having resolutely, steadily, and remarkably made a human-centric robot that actually works—and in demanding customer environments.” said Johnson. "AWS has revolutionized how we develop and train our robots, enabling us to scale our training infrastructure and accelerate innovation in autonomous robotics."