FeAR metric measures AI causal responsibility in continuous spaces
New metric assigns blame in robot-car interactions, scaling to real-world dynamics.
A team of researchers (Ashwin George, Luciano Cavalcante Siebert, David A. Abbink, Arkady Zgonnikov from TU Delft) has extended their Feasible Action-Space Reduction (FeAR) metric to continuous spatial interactions. Originally proposed for grid-world settings with discrete actions, FeAR quantifies causal responsibility by measuring how one agent’s actions constrain another’s feasible action set. In this new paper (arXiv:2505.17739v2), the authors formulate FeAR for continuous action spaces, making it applicable to real-world scenarios like autonomous vehicles and mobile robots sharing space with humans.
The metric works by comparing the actual action taken against the set of all possible actions that could have avoided a conflict or collision. For backward-looking responsibility, FeAR assigns blame proportionally based on how much each agent reduced the other’s feasible space. For forward-looking responsibility, it guides decision-making by estimating future responsibility. The researchers demonstrate FeAR on prototypical space-sharing conflicts (e.g., merging lanes, hallway crossings), showing that it captures nuanced causal dependencies missed by discrete models. The work is under review and has potential applications in AI safety, accountability in human-robot interaction, and designing autonomous systems that ethically share environments with people.
- Extends FeAR from discrete grid-world to continuous action spaces for real-world robot/vehicle interactions.
- Metric quantifies causal responsibility by measuring how one agent reduces another's feasible action space.
- Tested on prototypical spatial conflicts (merging, crossings) – applicable to backward blame and forward decision guidance.
Why It Matters
Assigns clear causal responsibility in autonomous systems, crucial for safety and regulatory compliance in human-shared environments.