100k Hours of Real-World Data Yields Only 57% Success in Robotics

eigenron · x · 2026-08-06

A recent test on the RoboCasa365 benchmark revealed that training with 100,000 hours of real-world trajectories (equivalent to 11.4 years) only achieved a 57% success rate. This stark data point has sparked reflections among practitioners about the current bottlenecks in embodied AI, suggesting the field is still in its infancy.

Original post →

More from Embodied

Embodied channel →