title: HakkoLab
emoji: 🧠
colorFrom: yellow
colorTo: pink
sdk: static
pinned: false
short_description: AI research lab of HakkoAI — companions & world models
The AI research lab of HakkoAI · emotionally intelligent companions today · world models & embodied agents next
Who we are
HakkoLab is the research arm of HakkoAI PTE LTD — the team behind HakkoAI (global) and 逗逗 / Doudou (China): one product, two brands. We build an AI character-companion app where people create characters and connect with them through text, voice and play, across iOS, Android and PC.
Our bet: the richest signal for the next generation of interactive AI doesn't live in static corpora — it lives in real, multi-turn, multimodal moments between people and the characters they care about. Building companions is how we learn what interactive intelligence actually requires — and HakkoLab exists to turn those lessons into research.
Two pillars
🫂 Product — companions people actually talk to. HakkoAI / 逗逗 puts an expressive, real-time AI companion in the pocket of users worldwide. Sustaining a believable presence — low latency, memory, emotion, multimodality — is a hard research problem in production, every day.
🧠 Research — from interaction to world models. We study how interactive, emotional, multimodal experience can be modeled, predicted and acted in — moving from companions on a screen toward agents that understand and operate in worlds.
Research interests
- World models — predictive models of interactive worlds (characters, scenes, dynamics) that can be rolled out and acted in, not merely described.
- Embodied agents — connecting world models to action, from in-app agents toward, eventually, physical embodiment.
- Learning from human–AI interaction — grounded, emotionally aware behaviour learned from how people and AI build rapport over time.
- Latent action representations — inferring "what just happened" between observations, so interaction without explicit action labels becomes usable for control.
- Multimodal, real-time companionship — speech, vision and text working together at conversational latency.
- Efficient & on-device inference — making all of the above run affordably at companion scale, on phones and PCs.
We draw inspiration from work such as WHAM (game-play → world model) and V-JEPA 2 (video → world model → robot control), and we're exploring how companion-scale interaction can push these directions further.
Research & releases
📄 world-model-research — our first public bundle: a survey of world-model & embodied-AI methods, latent-action notes, and a training / robotics-transfer plan. The starting point for everything above.
More model cards, datasets, demos and Spaces will follow. Follow HakkoLab to track what we publish.
Connect
🌐 hakko.ai · 📱 HakkoAI (Global) / 逗逗 (China) — iOS · Android · PC · 🤝 We're hiring — reach us through hakko.ai
中文
HakkoLab 是 HakkoAI PTE LTD 的 AI 研究实验室 —— 我们打造 HakkoAI（海外）与 逗逗 / Doudou（中国）：同一套产品、两个品牌的 AI 角色陪伴 App。用户在 iOS、Android 和 PC 上创建角色，并通过文字、语音与互动玩法与之建立连接。
我们相信：下一代交互式 AI 最宝贵的信号，不在静态语料里，而在人与所珍视的角色之间真实、多轮、多模态的互动之中。做陪伴产品，本身就是在学习"交互式智能到底需要什么" —— HakkoLab 的使命，就是把这些经验转化为研究。
两个支柱
- 🫂 产品 —— 真正被用户每天倾诉的陪伴。 HakkoAI / 逗逗 把一个富有表现力、实时的 AI 伙伴装进全球用户的口袋。低延迟、记忆、情感、多模态 —— 维系"可信的陪伴感"，是每天都在生产环境里被打磨的硬核研究问题。
- 🧠 研究 —— 从交互到世界模型。 我们研究如何对交互式、情感化、多模态的体验进行建模、预测，并在其中"采取行动" —— 让 AI 从屏幕里的陪伴，走向能理解并操作世界的智能体。
研究方向
- 世界模型 —— 可被推演、可被"行动"的交互世界预测模型（角色 / 场景 / 动态），而不只是被描述。
- 具身智能体 —— 把世界模型连接到行动，从应用内智能体，走向最终的物理具身。
- 从人机交互中学习 —— 从人与 AI 如何建立长期默契中，学到有据可循、有情感意识的行为。
- 潜在动作表征 —— 推断观测之间"刚刚发生了什么"，让没有显式动作标注的交互数据也能用于控制。
- 多模态实时陪伴 —— 语音、视觉与文本，在对话级延迟下协同工作。
- 高效 / 端侧推理 —— 让以上一切，在手机与 PC 上以陪伴级规模经济地跑起来。
我们从 WHAM（游戏数据 → 世界模型）与 V-JEPA 2（视频 → 世界模型 → 机器人控制）等工作中汲取灵感，并探索"陪伴级交互数据"能把这些方向推进多远。
研究与发布 📄 world-model-research —— 我们的首个公开 bundle：世界模型与具身智能方法综述、潜在动作笔记、训练 / 机器人迁移规划，是以上一切的起点。更多模型卡、数据集、Demo 与 Spaces 陆续上线。关注 HakkoLab，第一时间看到我们的发布。
🌐 hakko.ai · 📱 HakkoAI（海外）/ 逗逗（中国）—— iOS · Android · PC · 🤝 我们在招人 —— 通过 hakko.ai 联系我们