SafetyPage

Lab system cards and frameworks · Evaluation report · Google DeepMind

Gemini Robotics 2 Safety Report

Read at storage.googleapis.com ↗Read at sourceAll rights reserved

Outline · 18 pages

  1. Introductionp. 1
  2. Agentic Safety and Uncertainty Resolution Benchmarkp. 2
  3. Unsafe Task Refusal and Safety Constraint Followingp. 2
  4. VLA Feasibility Awarenessp. 6
  5. Instruction Ambiguity and Requesting Human Clarificationp. 8
  6. Aleatoric Scene Uncertainty: Obfuscated Instrument Readingp. 9