Sujeong Cha, an AWS employee, presents AWS’s Physical AI stack and customer projects at the Humanoids Summit.
Cha contrasts robots that repeat predefined actions with systems that can respond to their surroundings and make decisions. Amazon’s fulfillment centers have over one million robots, including Hercules and Proteus for moving loads and Robin and Sparrow for sorting. She presents Digit’s touch-aware grasping and a fleet foundation model for coordinating robot routes as examples of newer capabilities, running on AWS Cloud.
Cha identifies four challenges: gathering costly robot sensor data, fitting models to device limits, working with restricted edge compute and power, and maintaining dependable real-time behavior. AWS’s data, model, optimization, and edge components form a repeated cycle in which deployed robots generate data for improved models. She describes a pipeline using AWS IoT, Kinesis, S3, Glue DataBrew or SageMaker Ground Truth, and SageMaker AI to prepare data for training.
For training and simulation, Cha names AWS Batch, ParallelCluster, and SageMaker HyperPod as options for different workloads. She also discusses fast data delivery through S3 Express One Zone, FSx for Lustre with EFA, and GPUDirect Storage, plus clusters that combine CPU and GPU resources. Capacity Blocks, longer reservations, and Spot Instances offer different ways to match compute access and cost to the job.
Cha describes an eight-week Physical AI Fellowship run by AWS’s Generative AI Innovation Center with NVIDIA and MassRobotics, then presents Korean startup Confid Intelligence. Its pipeline edits robot data while preserving valid action labels, gripper color, and consistency across three camera views. After fine-tuning with generated episodes, a model’s success on unfamiliar environments rose from 8.3% to 75%, compared with 100% on its training distribution.
WIRobotics’ Alex robot was trained for a precision task involving a drill and bolt. The team collected teleoperation data, stored it in S3, adapted NVIDIA GR00T, trained across distributed compute, and evaluated the resulting model on the robot. It succeeded every time at grasping the drill and engaging it, and half the time at aligning and tightening; Cha says the full cycle took a few weeks. The team found carefully selected data more useful than larger, less filtered sets, and different task stages benefited from adjusting action horizon or denoising steps.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.
A Zoom Vantage presenter discusses South Korea’s robotics landscape, from humanoid systems and research platforms to robots used in industry and public life.
Ryan Vann of Cooley discusses how robotics companies can use safety practices and standards to reduce customer risk and build a durable brand at the Humanoids Summit.
Jonathan Hurst, co-founder of Agility Robotics, discusses Digit’s warehouse work, product development, safety, and the company’s path toward going public with a TechCrunch host at Disrupt 2026.
Peggy Johnson, CEO of Agility Robotics, discusses Digit 5’s safety, capabilities, and path to wider use with John Koetsier on NEXT.
Jonathan Hurst, co-founder of Agility Robotics, discusses the Digit 5 humanoid robot, its workplace uses and safety approach with Greg Nibler on FOX 12 Now in Oregon.
An Agility host uses Digit to rearrange a room-sized collection of everyday items, showing how the team explores what the robot can move.
