ARIS framework gives social robots relationship intelligence with RAG and knowledge graphs
Outperforms LLM baselines in perceived intelligence and likeability by significant margins
Deep Dive
Researchers have developed ARIS, an agentic AI framework for social robots that combines multimodal reasoning, a graph-based Social World Model, and retrieval-augmented generation (RAG). Tested on a Pepper robot in dyadic conversations, a user study (N=23) found ARIS scored significantly higher than an LLM baseline in perceived intelligence, animacy, anthropomorphism, and likeability. The system maintains bounded latency even after thousands of exchanges and will be open-sourced upon publication.
Key Points
- ARIS integrates a Social World Model — a knowledge graph that maps and updates relationships between users, enabling re-identification and social reasoning across multiple encounters.
- The RAG pipeline maintains bounded latency (response time) even after thousands of dialogue exchanges, solving a common scalability problem.
- In a user study with 23 participants using a Pepper robot, ARIS significantly outperformed a pure LLM baseline on perceived intelligence, animacy, anthropomorphism, and likeability.
Why It Matters
ARIS tackles a core barrier to social robots: retaining context and relationships, making long-term human-robot interaction truly viable.