Safe robot behavior requires more than obstacle avoidance. It requires understanding:
- what is in the scene
- what relationships matter
- what the current context implies for action
Does not depend only on the ROBOT!
The same action can be safe in one context but unsafe in another:
- moving a robot arm fast in free space vs. fast near a human hand
- pouring water into a cup vs. pouring near a keyboard or power outlet
Some safety constraints depend on object properties and hidden state:
- lifting a rigid closed box vs. lifting an open or deformable container with contents in it
Why not rely on scale alone?
- Long-tail failures: rare data
- No guarantee at execution time
Proposed Safety Taxonomy
