Stop Simulating, Start Moving: See It. Detect It. Act on It. The Intel Physical AI Challenge

AI is great on screen, but have you ever made it move? Welcome to the Intel Physical AI Challenge at Junction. This weekend, we’re taking your models out of the cloud and giving them hands, eyes, and real-time brains

Introduction

In many factories, people still inspect parts by eye and pull out the faulty ones, because programming a robot for the job takes specialist effort. With physical AI, you show a robot a task a few dozen times and it learns a policy it can run on its own. Ignite Next and Intel bring this to Junction, so you can take a real robot arm from first demonstration to an autonomous inspection cell in one weekend. You train your models on cloud GPUs and run them on Intel Core Ultra Series 3 (Panther Lake) laptops with Intel’s open-source robotics stack, which is built to take robot AI from the lab into production.

The Challenge

Each team gets its own station with an SO-101 robot arm, a leader arm for teleoperation, a top-view camera, an inspection camera and a box of Duplo bricks, some of which you mark as defective yourself.

Start by teleoperating the arm to record demonstrations, then fine-tune a vision-language-action (VLA) policy in Physical AI Studio on cloud GPUs. In parallel, train an Anomalib model that tells good bricks from defective ones at the inspection point. Then bring both together: the arm picks each brick, shows it to the inspection camera and drops it into the “good” or “defective” bowl, fully autonomously, with your models optimized with OpenVINO and running on the laptop’s CPU, GPU and NPU. Teams that get there early can push further with faster cycles, messier brick piles, new defect types or their own twist on the task.