Emotional Interaction Models
Emotional interaction models are AI systems that decide how a product should respond socially and emotionally, not only what factual answer it should give. In 我遇到了第一个真正想买的陪伴机器人！｜对话世博：越伴动力创始人【公路播客】, Yueban Dongli uses this idea for Xiaoban through an emotional brain, long-term memory, personality parameters, and generated action sequences.
这可能才是 AI 陪伴真正该有的样子｜对谈刷屏产品 EVE 创始人 Tristan adds a virtual companion version through EVE. Tristan says raw ChatGPT API calls felt too assistant-like, so Natural Selection / 自然选择 moved toward post-training on human companion-chat data and an emotional-interaction design team. The goal is not only warmer wording, but a model that can infer motive, relationship stage, remembered context, and response purpose before replying.
Key Points
- The model must map user speech, behavior, and household context into tone, posture, gaze, motion, and future relationship state.
- The episode distinguishes high-level decisions about what the robot should do from low-level action generation about how to execute the behavior.
- Shibo says companionship is not simple pleasing: a robot can appear sad, withdraw, stomp, turn away, or become less proactive toward a person who repeatedly mistreats it.
- Family World Simulator data is used to train and evaluate these interactions before enough real household data is available.
- On Device Fast Slow Brain supports emotional interaction by separating immediate behavior decisions from slower reasoning.
- EVE extends the concept from embodied behavior into text, voice, calls, 3D scenes, trust values, and character withdrawal.
- Human companion-chat data can become training material when the product goal is emotional value rather than factual task completion.
Connections
- Xiaoban and Yueban Dongli — product and company case.
- EVE, Tristan, and Natural Selection / 自然选择 — virtual companion case.
- Companion Robots and Robot Liveliness — product goals the models support.
- AI Companion Active Memory and AI Friend Products — memory and companion-product layer added by EVE.
- Qwen and Open Source AI Models — model base used in the source.
- World Models — adjacent modeling direction for predicting environment and interaction state.