Why blurry robot vision sometimes beats perfect sight

When language models control robots, giving them perfect sensor data backfires. Researchers tested LLM agents on a mechanical puzzle using three input types: raw camera images, depth-enhanced images, and perfect symbolic state. Counterintuitively, agents succeeded most often with noisy RGB video and failed most with ground truth. In simulation, deliberately corrupting action feedback improved performance 2.85× by disrupting repetitive loops where agents got stuck. The finding flips how we evaluate embodied AI: raw success rates hide whether models are truly reasoning or just compensating for sensor noise.