We are excited to share a new publication in Neural Computing and Applications:

Single-shot policy explanation to improve task performance via semantic reward coaching
by Aaquib Tabrez, Ryan Leonard, and Bradley Hayes.

In this work, we introduce Single-shot Policy Elicitation for Augmenting Rewards (SPEAR), a novel optimization algorithm that uses semantic natural language explanations to help humans and robots improve their decision-making. By modeling humans as reinforcement learning agents, SPEAR identifies misaligned reward functions and provides concise corrective feedback.

Highlights of the paper include:

• A formal characterization of the policy elicitation problem, framing behavior modification as semantic reward coaching.
• Introduction of SPEAR, which scales linearly with predicate count, substantially outperforming prior approaches.
• Human-subjects studies showing that SPEAR explanations improve policy comprehension, reduce cognitive load, and enhance decision-making.
• Demonstrations of robot-to-human and robot-to-robot collaboration, including evacuation scenarios and multi-agent tabletop cleaning tasks.

This work was supported by the Army Research Lab STRONG Program and highlights new opportunities for integrating explainable AI into human-robot teaming.

📄 Read the full paper here: DOI link CAIRO Link