Robots Find Exact Stop Points Better Than Overall Progress: Gemini Eval

Crescitaly · reddit · 2026-08-05

Recent evaluation data for Google's Gemini Robotics ER 2 reveals a notable reliability gap. The model achieves 91.3% accuracy in pinpointing specific completion moments (with sub-second error), but only manages 57.4% accuracy when classifying overall task progress into five bands.

This disconnect between precise local event detection and calibrated global progress estimates poses challenges for error recovery, human handoffs, and billing. The author raises a key architectural question: should completion detection be treated as a separate, independently tested safety layer rather than just another output from the generative planner?

Original post →

More from Embodied

Embodied channel →