Manufacturing 50,000 Humanoid Robots Next Year | Bernt Børnich, 1X

In this talk

1. 0:00

   Scaling production while protecting quality

   Børnich says 1X’s Hayward factory is designed for 10,000 robots a year, while a San Carlos facility is being built for 100,000 at full capacity. The company aims to ship 50,000 next year, but plans to increase volume through quality gates because rare defects emerge as production grows. He says some deployments will need structured environments, since placing robots into thousands of unrelated tasks would be difficult to manage.

2. 3:52

   Learning from homes and closing the data loop

   After about three years of home pilots with the wheeled Eve robot, Børnich says progress has been slow because each hardware change requires new deployments and data. He describes expert operators as a way to test tasks before automating them, and says 1X is pursuing a general model trained on diverse human and robot data. A larger fleet could supply more real-world experience, but he says diversity of tasks and environments matters more than repeating the same task many times.

3. 9:15

   Designing robots for faster iteration

   Børnich says 1X developed its own motors and tendon drives, which he describes as simpler than systems built around complex gear trains, though they require in-house manufacturing. A major system change can take about four weeks to reach a walking robot on the line, giving the team a way to address assembly and quality issues. He also describes reducing cables, connectors, and parts, and says he would have kept the product team smaller for longer.

4. 20:15

   Building trust, warmth, and useful behavior

   Børnich says 1X aims for robots that blend into society through warm materials and approachable design. He expects models trained on broad data to learn social behaviors, such as handing over an object and releasing it when someone takes hold. He also sees play and exploration as potential ways for robots to learn about their surroundings, while acknowledging that shaping desirable behavior from imperfect data remains an open problem.

5. 31:14

   Choosing applications and enjoying the work

   Børnich expects a few general models to benefit from large datasets, with specialized models serving narrower tasks in the meantime. He says next year’s deployments are expected to include enterprise customers, homes, and the Neo platform, while 1X is not planning to make clothes folding its main application. He describes startup work as a continuing sequence of problems and says founders need to enjoy the process.