The iCub storytelling system achieved 88% trial success and 86% VLM sticker-description agreement, with the LLM adding to human contributions in 8% of trials and never correcting them¶

Claim · All claims
Evidence · 1 study (2 entries) · design r2 · q3
Currency · newest study 2025 · not yet checked against its sources · not reviewed by a person · recheck by 2028: tool and model accuracy changes quickly

Subclaims¶

q3 i? Across the validation trials the system was robust with an 88% success rate despite failures such as wrong sticker detection, microphone issues and LLM crashes. → Maria Lombardi 2025 q3 i? Zero-shot VLM sticker descriptions agreed with the participant's chosen image in 86% of cases, and the LLM added a small contribution to the human's story piece in 8% of trials while never correcting or polishing it. → Maria Lombardi 2025 (2)

Evidence¶

Maria Lombardi 2025¶

Maria Lombardi, Carmela Calabrese, Davide Ghiglino, Caterina Foglino, Davide De Tommaso, Giulia Da Lisca, Lorenzo Natale, Agnieszka Wykowska. (2025). Would you let a humanoid play storytelling with your child? A usability study on LLM-powered narrative Human-Robot Interaction. https://arxiv.org/abs/2508.02505 doi:10.48550/arxiv.2508.02505

q3 · i? · design · r2

Quantitative failure analysis of the validation experiment, logging FSM transitions and trial success across the many interacting modules. The article reports an overall 88% success rate; the 12% failed tests are analysed in the Limitations section.

"Our system was pretty robust with 88% success rate."

Maria Lombardi 2025 (2)¶

Maria Lombardi, Carmela Calabrese, Davide Ghiglino, Caterina Foglino, Davide De Tommaso, Giulia Da Lisca, Lorenzo Natale, Agnieszka Wykowska. (2025). Would you let a humanoid play storytelling with your child? A usability study on LLM-powered narrative Human-Robot Interaction. https://arxiv.org/abs/2508.02505

q3 · i? · design · r2

VLM detection evaluation computed as agreement between the LLM's sticker description and the image chosen by the participant, yielding 86% overall agreement. A separate LLM-generation analysis found additions in 8% of trials and no corrections of the human's contribution.

"The overall agreement was 86%."

Discussion¶

Related Claims¶

• Adults rated the LLM-powered iCub storytelling system as usable, with a total SUS score of 76.70 indicating good product usability — related
• Participants liked the co-created stories (mean likability 8.61) and perceived them as coherent (mean coherence 8.28) on a 1-10 scale — related