safety filters for competitive play
turning a fault rule into the geometry of the strategy space
Two quadrupeds play a zero-sum interaction on a 5.2 × 3.0 m field. The safety specification is coupled: whether one robot can stay safe depends on what the other does. Instead of paying for safety with a reward penalty, the rule is encoded as a least-restrictive filter around the policy, so it reshapes the set of admissible strategies rather than the score.
Add figures with the three-column grid:
Replace this file, and this caption, with a real render.