3 September 2026

Robotics startups experiment with methods to gather training data

• Companies including Shift and Scale AI are testing different ways to collect demonstration data, from video recording to having humans directly operate robots.
• Some startups use exoskeletons or special gloves to capture human movements in formats that robots can learn from.
• The field faces a data shortage, with the largest public robotics dataset containing only 3,500 hours of demonstrations, below what companies believe they need.

How it was covered

Understanding AITimothy B. Lee

Companies like Shift, Scale AI, and others are experimenting with various data collection methods including recording human actions on video, hiring humans to directly operate robots, and using exoskeletons or special gloves to capture human movements in robot-compatible formats. The newsletter emphasises that solving the data shortage is critical because the largest publicly available robotics dataset has only 3,500 hours of demonstrations, far below what industry expects is needed.