Why Imitation Learning Needs Better Data: The Annotation Quality Gap

Imitation learning has emerged as a powerful paradigm for training robotic foundation models, but it suffers from a critical bottleneck: data quality. Traditional computer vision datasets are often too noisy, lack high-fidelity pose tracking, or miss the subtle contact-rich interactions required for manipulation tasks.

At Radioception, our research shows that increasing annotation quality from 80% to 95% confidence can yield a 3x improvement in policy success rates. The gap lies in capturing not just what the human did, but the physical forces, precision, and intent behind the action.