2025 · Google DeepMind · Advanced
ASIMOV Benchmark for Robot Semantic Safety
ASIMOV evaluates whether robot-brain foundation models understand unsafe physical situations and safety rules.
Direct answer
What does ASIMOV Benchmark for Robot Semantic Safety contribute?
ASIMOV evaluates whether robot-brain foundation models understand unsafe physical situations and safety rules.
Background
The benchmark turns real-world injury narratives and operational constraints into multimodal safety evaluations for embodied AI systems.
Problem
The work addresses a central constraint in Safety: building systems that learn useful representations or actions while remaining general enough to transfer beyond a single demonstration or environment.
Core idea
ASIMOV evaluates whether robot-brain foundation models understand unsafe physical situations and safety rules.
Architecture and method
The benchmark turns real-world injury narratives and operational constraints into multimodal safety evaluations for embodied AI systems.
- Semantic safety dataset collection
- Robot constitution generation
- Multimodal unsafe-situation evaluation
Results and impact
Safety cannot be added after deployment; ASIMOV is useful because it tests semantic risk before an action reaches a robot body.
Prerequisites
- Safety
- Embodied reasoning
- Benchmark design
Recommended reading order
Read the explanation above, review the related topic pages, then use the primary-source links below to inspect the abstract, figures, experiments, and released implementation.
Primary sources
External links are provided after the context needed to evaluate the work.
Follow-up research
Related papers and concepts
RT-2: Vision-Language-Action Models Transfer Web Knowledge to Robotic Control
RT-2 co-trains vision-language models on web and robot data so semantic knowledge can influence actions.
GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
GR00T N1 uses a dual-system architecture for language reasoning and continuous humanoid control.
Gemini Robotics-ER 1.6
Gemini Robotics-ER 1.6 improves spatial reasoning, multi-view understanding, tool use, and safety-oriented robot reasoning.
OpenVLA: An Open-Source Vision-Language-Action Model
OpenVLA is an open 7B-parameter VLA trained on the Open X-Embodiment dataset.
Common questions
Frequently asked questions
What is the main idea of ASIMOV Benchmark for Robot Semantic Safety?
ASIMOV evaluates whether robot-brain foundation models understand unsafe physical situations and safety rules.
Why is ASIMOV Benchmark for Robot Semantic Safety important?
Safety cannot be added after deployment; ASIMOV is useful because it tests semantic risk before an action reaches a robot body.
What should I learn before reading ASIMOV Benchmark for Robot Semantic Safety?
Recommended prerequisites are Safety, Embodied reasoning, Benchmark design.