Georgia Tech PhD student trains humanoid robots with AR glasses

Call it magnificent mimicry.

The rapid advancement in humanoid robotics has been limited by slow, manual data collection methods requiring direct robot operation. Georgia Tech researchers have developed a breakthrough approach using Meta‘s Project Aria glasses to capture human behaviors that can train robots more efficiently.

Key innovation: EgoMimic, developed by PhD student Simar Kareer at Georgia Tech’s Robotic Learning and Reasoning Lab, uses egocentric recordings from Aria glasses to create training data for humanoid robots.

• The framework combines human-recorded data with robot data to teach robots everyday tasks
• Traditional robot training requires hundreds of manual demonstrations through direct robot control
• EgoMimic achieved a 400% performance improvement across various tasks using just 90 minutes of Aria recordings

Technical implementation: Project Aria glasses serve dual purposes in the research, functioning both as a data collection tool and as the robot’s visual system.

• The glasses are mounted on humanoid robots to provide real-time environmental perception
• Aria’s Client SDK streams sensor data directly to the robot’s control system
• Using identical hardware for both human demonstration and robot operation reduces the “domain gap” between training and execution

Research implications: The breakthrough could enable more efficient and scalable robot training methods.

• Traditional robot training requires task-specific demonstration data that is costly and time-consuming to collect
• EgoMimic can leverage existing datasets like Ego4D, which contains over 3,000 hours of human activity recordings
• The system successfully performed tasks even in previously unseen environments

Industry perspective: Meta sees Project Aria as a catalyst for collaborative robotics research.

• James Fort, Reality Labs Research Product Manager, emphasizes the importance of standardization in egocentric research
• The technology could enable broader collaboration among researchers
• The research will be presented at the 2025 IEEE International Conference on Robotics and Automation

Future implications: This research represents a significant step toward more capable and adaptable humanoid robots that could transform how we approach everyday tasks, though questions remain about real-world scalability and the breadth of tasks that can be effectively learned through egocentric data alone.