Robotics founder: Error recovery is the true test of an AI

This article summarizes an episode of Ryan Peterman’s video series featuring Sergey Levine, co-founder of Physical Intelligence.

Sergey Levine, co-founder of Physical Intelligence and a professor at UC Berkeley / Photo credit: UC Berkeley Research

Perfect robot videos hide the real facts investors need to measure AI progress. Sergey Levine, co-founder of Physical Intelligence and a professor at UC Berkeley, says true machine intelligence comes from handling unexpected situations over and over.

This means companies must focus on real-world fixes rather than edited marketing videos.

Flawless videos mask brittle robotics

Evaluating commercial readiness requires stepping away from controlled demonstrations to track how systems handle unexpected failures.

A basic task executed in a novel environment carries weight if it proves the software can recalibrate itself, requiring executives to look for specific markers:

• Evaluate repeated trials. Verify consistent performance rather than one-off successes.
• Monitor error recovery. Watch how the system behaves immediately after making a physical mistake.
• Identify autonomous learning. Ensure the machine updates its own behavior in new environments without manual reprogramming.

As Levine puts it,”generalization often doesn’t look that impressive when viewed in isolation because it is a property of many trials, not one trial.”

Diverse environments drive AI education

Moving beyond isolated testing requires feeding systems a diverse diet of real-world experiences to capture learning opportunities:

• Educational deployments: View every customer location as a chance to teach the system new, transferable skills.
• Task variety: Maintain a wide range of distinct actions even during initial hardware launches.
• Messy environments: Progress from organized safety facilities into unpredictable settings to maximize adaptability.

Repetitive loops harm long-term development. As Levine explains, “data is more like an education program for your robot than a fungible commodity.”

Embodied data precedes digital training