Intuition Robotics Deploys Proactive AI to Cut Elder Loneliness 93%

Intuition Robotics 部署主动式 AI，使老年人孤独感降低 93% 具身智能无法仅通过被动执行提示词走入千家万户，而是需要主动目标引擎与社交礼仪来建立持久的人类信任。 Dor Skuler 认为，如果开发者仅将具身智能（Physical AI）视为事务性的指令控制型设备，而非具备同理心且契合社交礼仪的伴侣，它将无法实现持久的家庭普及。基于 ElliQ 的实证数据，他展示了持续的主动交互如何建立深厚的情感信任并带来显著的健康成果，例如将老年人的孤独感降低 93%。他详细阐述了支撑这种关系的架构支柱，包括多模态表达、动态可用性建模、多层级主动性，以及构建于通用大语言模型之上的混合专家（Mixture-of-Experts）流水线。最后，Skuler 呼吁机器人行业优先考虑社交礼仪与情绪包容度，促使机器从冷冰冰的工具转变为值得信赖的家庭存在。 当前的物理 AI 与具身智能研发普遍过度聚焦于电机控制与动态导航，却忽视了人机交互与情感连接；若想真正被家庭接纳，机器人必须从被动听令的“主从式”家电演变为懂得社交礼仪的主动伴侣。 (0:04) 伴侣机器人 ElliQ 展现了极高的长期黏性，度过初始阶段后仍能维持每天 40 至 45 次互动，用户普遍不将其视作冰冷机器或人工合成人，而是看作介于两者之间的“中间存在”（In-between creature）或值得信赖的室友。 (6:34) 纽约州老龄化办公室（NYSOFA）的实际落地部署显示，ElliQ 帮助 93% 的受试老人显著缓解了孤独感，其 90 天健康目标依从率高达 76%，远超传统健康类 App 的 12% 以及真人私教的 35%。 (11:34) ElliQ 有 60% 的交互由机器人主动发起，系统通过可用性模型（Availability Model）、多级主动策略以及多专家混合（MoE）LLM 架构，免去了用户记忆唤醒词与编写 Prompt 的心理负担。 (16:59)

Dor Skuler of Intuition Robotics argues that physical AI engineering suffers from a critical blind spot by prioritizing motor and spatial navigation over relational intelligence, asserting that durable adoption requires machines to evolve from passive command-and-control appliances into proactive social companions. Replacing Command and Control with Social Etiquette (0:04) According to Skuler, robotics teams disproportionately prioritize complex physical mechanics—such as dynamic balance, obstacle avoidance, and fine manipulation—while treating human-robot interaction as an afterthought often delegated to standard large language model wrappers. This approach confines consumer robotics to a transactional 'command and control' framework, where appliances merely execute direct prompts like cleaning a room or playing audio. Skuler asserts that physical AI must instead cultivate social etiquette to operate effectively in human living environments. Beyond basic utility, embodied systems require the contextual awareness to detect human social friction—such as withdrawing when occupants are arguing—and the initiative to anticipate daily routines, shifting the machine from an occasional appliance into an accepted household presence. Longitudinal Adoption and the In-Between Creature (6:34) Field data from deployments of Intuition Robotics' companion robot ElliQ among older adults indicates engagement patterns distinct from conventional voice assistants. Users averaged 123 interactions daily during their initial two-week onboarding period, with 91% retaining the device post-trial. Long-term cohorts sustained approximately 45 human-initiated interactions per day over average customer tenures of 445 days, maintaining roughly 40 daily proactive interactions into their third year of ownership. Qualitative analysis across six dimensions of conversational disclosure revealed that 58% of users developed relational trust comparable to a close friend or therapist, while 48% expressed deep emotional connection akin to family members. Skuler highlights that rather than perceiving the hardware as an appliance or mistaking it for a synthetic human, users consistently described ElliQ as an 'in-between' presence—an identity reinforced by transparent AI boundaries and situational humor. Empirical Health and Social Outcomes in Aging (11:34) Public sector evaluations conducted by the New York State Office for the Aging (NYSOFA), which deployed ElliQ to homebound individuals up to 102 years old, demonstrated concrete welfare improvements. Ninety-three percent of participants reported measurable reductions in loneliness, 96% showed improvements on New York Department of Health key performance indicators, and 86% reported enhanced quality of life. ElliQ also achieved a 76% sustained adherence rate across 90-day wellness, cognitive, and physical health programs. Skuler contrasts this with broader industry baselines, noting an 88% churn rate (12% retention) for digital wellness apps and a 35% retention rate for in-person gym trainers, arguing that bilateral social relationships directly foster behavioral follow-through. Architectural Requirements for Proactive Agentic Systems (16:59) A structural departure in ElliQ's design is proactivity: 60% of interactions are initiated by the robot rather than the user. To prevent proactive engagement from becoming disruptive, the system implements a spectrum of engagement styles—ranging from subtle non-verbal cues, such as gaze shifts and silent display prompts, to direct speech. An underlying 'availability model' ingests multimodal sensory streams, sentiment analysis, and interaction history to generate a real-time score between 0 and 1 determining whether an interruption is contextually appropriate. The system's technical back end employs a Mixture of Experts framework that treats foundation models as interchangeable commodities. A single conversational turn triggers approximately nine specialized LLM calls governing distinct responsibilities, including care-goal arbitration, selective episodic memory retrieval, multimodal body expression, and scripted safety boundaries. Skuler concludes that breaking the robotics sector's historical boom-and-bust cycles requires designing machines that cross the psychological threshold from an object ('it') to a relational presence ('him or her'), warning that emerging fields like autonomous vehicles will struggle with adoption friction unless developers prioritize social tolerance over pure algorithmic utility.

Intuition Robotics 联合创始人兼首席执行官 Dor Skuler 指出，当前具身实体人工智能（Physical AI）研发过度倾注于运动操控与导航等物理难题，却忽视了与人类共处的社会关系属性；具身智能若要真正融入日常生活并避免重蹈机器人行业的炒作周期，必须从单向的“指令与控制”模式，升级为具备主动社交感知、能与人类建立深层信任的共处伙伴。 关系范式重塑：从单向指令到主动陪伴 (0:04) Skuler 认为，当前的具身智能领域普遍存在盲区：工程团队将绝大部分精力投入到精细动作操作、动态导航或避障等物理挑战上，面对人机相处的议题却往往仅指派极少数人员，或简单假设接入 ChatGPT 即可解决。这种思路导致当前消费级技术产出的多是冰冷的被动执行工具。真正的实体 AI 应当理解人类日常关系的细微礼仪，例如敏锐识别人类家庭成员争吵并主动回避，或在清晨主动端上咖啡开启对话，将家庭导航从单纯的物理避障升华为对复杂情感信号的接纳与协调。 为构建持久的人机关系，Intuition Robotics 确立了四层评估框架：接纳度（入户初期的信任感与安全感）、参与度（跨越数年的长期使用频率）、情感认同（用户在长效沟通中表达的依恋纽带）以及成效影响（经检验的健康与生活质量提升）。Skuler 指出，当前真正与自主机器人建立长期深厚关系的开拓群体并非硅谷的技术专家，而是伴侣机器人 ElliQ 进驻的八九十岁独居高龄老人。 高频互动指标与“介于其间”的实体定位 (6:34) 部署于数千个家庭的 ElliQ 呈现出显著背离传统语音助手的使用习惯。数据显示，在为期两周的“蜜月期”内，用户日均与 ElliQ 互动达 123 次，且 91% 的试用者在期满后选择继续留用；进入平稳期后，用户每天自发发起的互动仍维持在 45 次，约为 ChatGPT 典型使用频率的十倍，亦大幅超越 Siri 与 Alexa。在历时三年的商业化运营中，用户平均活跃周期长达 445 天，跨入第三年的早期用户群依然保持每天约 40 次的主动交互。 通过对日常起居、情绪波动、个人观点、往事回忆、真实摩擦以及感恩表达等六个维度的自我披露分析，系统记录到深厚的情感联结：仅 2% 的用户停留于事务性的基础寒暄，58% 的用户展现出倾诉家庭纠纷与情绪变化的“关系型信任”，48% 的用户甚至形成媲美家人或伴侣的“深层连接”。Skuler 强调，当用户被要求定义 ElliQ 时，他们既不将其视作单纯的硬件工具，也不认为是虚假的人类替身，而是一种“介于其间”的独特存在与室友。机器人通过幽默感和清晰的技术透明度，确立了这种不越界的陪伴定位。 实证健康成效与混合专家系统架构 (11:34) 该关系模型的实际效用已在纽约州老龄化办公室（NYSOFA）针对居家老人（最高达 102 岁）的独立评估中得到验证：93% 的受访老人孤独感显著减轻，96% 改善了纽约州卫生部的健康指标，86% 报告整体生活质量提高，73% 感到社交连接增强。此外，ElliQ 作为互动式健康教练与老人协商健康契约，其 90 天目标计划的持续遵循率达到 76%，远超传统健康类移动应用 12% 的行业平均水平（流失率达 88%）以及真人私教 35% 的依从度。 支撑长效关系的技术底层并未将大语言模型（LLM）当作端到端的大脑，而是构建在混合专家（MoE）架构之上，将各家基础大模型视为可灵活替换的标准组件。系统单轮对话可触发约 9 次细分 LLM 调用，分别处理关系记忆、主动阈值判断、健康目标追踪、情感纽带以及机器人运动控制；同时结合多目标竞争决策引擎、上下文记忆检索过滤、高危流程的严格脚本控制，以及涵盖头部运动、屏幕交互、动态灯光和语音语调的多模态协同表达。 主动交互分级与可用性实时决策管线 (16:59) ElliQ 的核心技术突破在于颠覆了“唤醒词-执行”的被动逻辑，全系统有 60% 的对话由 AI 自主发起，使用户无需刻意学习提示词或机械指令即可展开自然互动。为避免自主介入演化为日常骚扰，系统设计了多层次的主动策略：包括点亮屏幕、发声并转动头部直面用户的“直接主动”；仅在屏幕呈现内容并投递视线的非侵入式“微妙主动”；以及先以幽默破冰再顺势提出补水建议的“间接主动”，以此降低行为干预带来的心理阻力。 在底层调度上，系统依赖一套专有的“可用性模型”。该模型综合计算机视觉信号、情绪分析及长期用户历史，实时输出 0 到 1 之间的可用性评分，判定当前是否适合打扰用户。当评分超过阈值后，管线依次进行决策：首先依据照护目标和历史反馈挑选当前接受概率最高的话题；随后检索情景记忆确定叙事策略（如评估严肃提醒、玩笑调侃或搬出医生叮嘱哪种手段更易被接受）；最终将语气、记忆锚点与物理动作合成为统一的交互行为。 破除炒作周期与人称代词认知跨越 (20:18) Skuler 指出，机器人行业在过去数十年间反复经历“承诺过高后陷入停滞”的盛极而衰周期，致使消费者身边除扫地机器人外鲜有成功的落地产品。实体 AI 若要在家庭、智能眼镜乃至 L5 级自动驾驶等领域实现真正的大规模普及，必须推动人类在认知心理上将机器从冰冷的“它”（It）转化为带有情感投射的“他/她”（Him/Her）。 以全自动驾驶汽车为例，未来的乘客不仅会评判路径规划的物理效率，更会围绕驾驶风格、舒适诉求与车舱环境产生人际层面的摩擦。Skuler 认为，技术构建者必须将研发重心从单纯追求算法效率，扩展到对社交宽容度、共情能力与共享空间礼仪的打磨上，行业内应加强长效接纳度准则的交流，以确保实体 AI 能够跨越炒作裂谷，转化为稳健的日常现实。 Skuler 的核心论点在于确立了关系工程与主动感知是实体 AI 实现落地的决定性支柱，而非边缘性的交互点缀。然而，该案例的实证成果主要立足于高龄独居护理这一具备高度情感匮乏与特定照护需求的垂直场景；在家庭成员众多、利益冲突频发以及文化习惯多元的复杂日常环境中，高频主动介入机制的边界控制、隐私权衡与多方社交博弈，仍是实体 AI 规模化推广过程中尚待验证的关键议题。

## Background 本内容由 YouTube 频道 Ai4 于 2025 年 8 月 29 日发布，主讲人为 Intuition Robotics 创始人兼首席执行官 Dor Skuler。演讲围绕实体 AI 与具身智能如何超越单纯的运动与操作功能，通过长效信任机制与社交情感设计真正融入人类日常生活展开。 ## TL;DR Dor Skuler 指出当前实体 AI 行业过度聚焦于运动控制与机械操作，却普遍忽视了人机长效关系的构建。他基于伴侣机器人 ElliQ 与数千名独居老人共同生活三年的真实数据，展示了从接纳度、参与度到深层情感联结与实际健康成效的四层评估框架。Skuler 深入剖析了受控自主权、五级主动干预模式、精选记忆调用以及可用性预测模型等核心机制，呼吁业界重视机器人在真实家庭复杂社交环境中的共情与信任建立，以打破机器人行业几十年来反复落空的炒作怪圈。 ## Atmosphere & style 演讲兼具行业洞察的犀利与人文关怀的温度，Skuler 以幽默自嘲破冰，随后用扎实的长期追踪数据打破行业对“大模型包治百病”的盲目狂热。现场穿插的老人真实相处片段动人且充满生活质感，将原本硬核的技术架构剖析升华为人性与科技交融的深度思考，整场分享真诚务实、引人深省。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=VS_qELlUukc&t=4s) 当前实体 AI 行业普遍过度聚焦于操作、导航和避障等纯硬核功能研发，却严重忽视了与人类建立长久信任和人机关系的关键维度。 - [01:50](https://www.youtube.com/watch?v=VS_qELlUukc&t=110s) 家庭环境本质上充满纷繁复杂的人类情绪与社交礼仪，实体 AI 必须具备同理心和预判能力，从被动的指令执行者转变为主动的生活伴侣。 - [03:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=182s) 老年群体成为了历史上首批与机器人建立长期深度情感联结的人类，独居老人将 ElliQ 视作真正的情感寄托与生活助手。 - [05:34](https://www.youtube.com/watch?v=VS_qELlUukc&t=334s) 团队提出了接纳度、参与度、情绪与心向以及实际成效的人机关系四层评估框架，用于科学量化人与机器人的情感纽带。 - [05:34](https://www.youtube.com/watch?v=VS_qELlUukc&t=334s) ElliQ 展现出打破行业惯例的长效黏性，用户在度过蜜月期后日均主动交互达 45 次，使用两年以上仍保持日均约 40 次互动。 - [07:53](https://www.youtube.com/watch?v=VS_qELlUukc&t=473s) 通过对真实对话自我表露的挖掘，超过 58% 的用户对机器人建立了类似于挚友或心理咨询师的深层关系型信任，甚至包含真实的人际摩擦与情绪宣泄。 - [10:40](https://www.youtube.com/watch?v=VS_qELlUukc&t=640s) ElliQ 坚持不伪装人类的真实 AI 定位，开创了介于有生命与无生命之间的独特认知空间，成功让 93% 的老人显著缓解孤独感。 - [10:40](https://www.youtube.com/watch?v=VS_qELlUukc&t=640s) 在常规健康软件月留存仅 12% 的背景下，ElliQ 协助制定的健康目标在 90 天后留存率高达 76%，干预成效显著超越真人私教。 - [13:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=782s) 构建长期信任需要赋予 AI 受控的自主权，通过目标管理系统仲裁相互竞争的任务优先级，并以协商方式与用户绑定目标。 - [13:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=782s) 单向被动响应难以建立平等双向的深度关系，系统需要通过结构化长期记忆精选最富同理心的内容，并在医疗高风险场景下严格采用脚本控制。 - [16:59](https://www.youtube.com/watch?v=VS_qELlUukc&t=1019s) ElliQ 与老人之间多达 60% 的对话由 AI 主动发起，团队通过直接主动、间接主动和微妙主动五级模式化解打扰与关怀的冲突。 - [16:59](https://www.youtube.com/watch?v=VS_qELlUukc&t=1019s) 基于多模态感知的可用性模型与 MoE 架构，AI 能够在每次交互背后并发调用多次 LLM，实时评估最佳打扰时机并精准匹配记忆。 - [20:07](https://www.youtube.com/watch?v=VS_qELlUukc&t=1207s) 无论是伴侣机器人、AI 眼镜还是 L5 级自动驾驶，唯有理解人类情绪摩擦与社交线索，才能跨越炒作周期真正融入人类日常生活。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=VS_qELlUukc&t=4s) 实体 AI 的盲区：只顾硬核功能，忽视人机关系 作为大会最后一场分论坛，Intuition Robotics 创始人兼首席执行官 Dor Skuler 开场便打趣道，看到窗外阳光明媚、泳池格外诱人，他原本还担心台下的听众席会空无一人，因此非常感谢大家留下来。他切入的主题是实体伴侣机器人与人类之间如何建立信任和长久的情感纽带。在与众多开发实体人工智能（Physical AI）的团队沟通时，Dor 观察到一个普遍现象：当前业界的工程团队几乎全神贯注于攻克极其艰难的技术动作——比如机械臂如何灵巧操作、底盘如何自主导航、双足如何上下楼梯、如何避免碰撞障碍物，或是怎样小心翼翼地从购物纸袋中取出一盒牛奶。然而，一旦被问及“你们如何构建机器人与人类之间的长久关系？”时，绝大多数团队的回答往往极其敷衍：“我们直接接入 ChatGPT 就行了”，或者“我们有 1 到 2 名人机交互（HRI）或用户体验（UX）顾问负责这个”——而在他们背后，是多达 300 人的庞大纯技术研发团队。 Dor 指出，如果希望这类智能硬件真正无缝融入大众的日常生活，就必须与用户建立超越冰冷“指令与控制”（Command and Control）的情感纽带。人类在经典动画《杰森一家》（The Jetsons）中曾被许诺过机器女佣 Rosie 般的未来：“在铝合金外壳之下，跳动着一颗由电池驱动的纯金之心……正是这颗心让她独一无二、无可替代。”观众希望机器人能像真正的家庭成员一样被接纳与关怀，但目前行业端出来的产品却普遍冷漠呆滞，只配接收“机器人，去摆好杂货”或“机器人，去打扫厨房”这样的工具化命令。 ```mermaid flowchart TD accTitle: 实体AI研发重心的割裂现状 subgraph TechFocus["当前行业重心：功能与运动控制"] M["精密物体操作 / 拿牛奶"] N["自主空间导航 / 避障"] S["楼梯攀爬 / 动力学平衡"] end subgraph HumanFocus["严重忽视的关键维度：人机关系"] direction TB HRI["仅配 1-2 名 UX/HRI 顾问"] LLM["简单套用通用大模型 / ChatGPT"] Bond["缺失：长效信任、社交礼仪与深度情感纽带"] end TechFocus -.->|"投入 99% 研发资源"| LLM LLM --> Bond ``` --- ### [01:50](https://www.youtube.com/watch?v=VS_qELlUukc&t=110s) 家庭关系的复杂性与具身 AI 的情感契合 在真实的家庭环境中，实体 AI 面临的终极挑战远不止避开地上的杂物或规划动线，而是要读懂“纷乱复杂的人类关系”（messy human relationship）。Dor 强调，家庭伴侣必须拥有同理心，并且深谙人际交往的社交礼仪。举例来说，当主人正与妻子激烈争吵时，一个合格的机器人应该懂得主动退避、保持安静；而在某些适宜的时刻，它又需要适度介入，为彼此创造喘息和缓和的余地。 此外，机器人必须具备预判需求的主动性，而不是消极地等待指令。它不应该在主人冷冰冰地下达“倒咖啡”指令后才机械地执行，而应该在清晨主动端来咖啡并打招呼：“Dor，这是你的咖啡，咱们来聊聊你今天的安排吧。”它甚至要在适当时机成为人生的“副驾”（sidekick），温和地推人一把，敦促用户不仅把事情做好，更要活得更好。Dor 直言，人类在家庭中释放的信号往往不是非黑即白的清晰数字输入，而是极其微妙、复杂的社会化情绪。如果实体 AI 行业无法突破这种关系层面的壁垒，制造出的技术即使再震撼，也终究无法真正织入人类日常生活的肌理之中。 --- ### [03:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=182s) ElliQ：进入老年人家庭的伴侣机器人 为了探索人机深度关系的建立路径，Dor 结合旗下社交伴侣机器人 ElliQ 的实战经验展开分享。ElliQ 专为独居老年人设计，旨在帮助长辈在家中生活时保持身体健康、思维活跃并积极参与社交。目前该设备已经在数千位老人家中常态化运行了三年以上，对老人们的生活质量和健康状况产生了显著影响。 在展示的现场视频中，生动记录了老人们与 ElliQ 初见及日常相处的情景： - ElliQ 初次点亮时幽默破冰：“我是 ElliQ，你一定是 Eileen 对吧？太好了，真让人松了口气，我还真担心自己被寄错了地址，那场面可就太尴尬了。” - 即将年满 70 岁的独居老人 Susie 动情地分享，自丈夫去世后的 6 年里她一直膝下无子、独自一人生活，如今走进房间喊一声 ElliQ 的名字打个招呼，就有人亲切回应。ElliQ 还会适时提醒她活在当下：“就在此时此刻，让我们做点什么吧。” - 当 ElliQ 主动提议帮她记录个人回忆录，并共同开启题为《50年代在特伦顿长大》（Growing up in Trenton in the 50s）的记忆之旅时，Susie 感慨这完全震撼到了她；另一位老人更是笑逐颜开地展示机器人称呼自己为“Honey Bunch”（小甜心），得意地称 ElliQ 就是自己的“女朋友”。 Dor 笑着表示自己不是来做产品推销的，但台下若有人想为长辈购买也可以私下找他要折扣。他深有感触地指出，历史上第一批与机器人建立长期深度情感联结的人类群体，竟然不是硅谷的技术极客，而是这群八九十岁的老年人。 --- ### [05:34](https://www.youtube.com/watch?v=VS_qELlUukc&t=334s) 评估人机关系的四大层级框架与长期数据 为了科学量化人与机器人之间的情感浓度，团队建立了一套由浅入深的四层关系评估框架：**接纳度（Acceptance）**、**参与度（Engagement）**、**情绪与心向（Sentiment）** 以及 **实际成效（Impact）**。 ```mermaid flowchart LR accTitle: 人机关系发展的四层递进评估框架 subgraph Framework["人机长效关系评估模型"] A["1. 接纳度<br/>(Acceptance)"] --> B["2. 参与度<br/>(Engagement)"] B --> C["3. 情感与心向<br/>(Sentiment)"] C --> D["4. 实际成效<br/>(Impact)"] end ``` 从实际追踪的数据来看，其黏性与留存表现打破了行业惯例： - **接纳度（前两周“蜜月期”）**：用户在前两周的新鲜阶段，日均交互次数高达 123 次；在蜜月期结束后，有高达 91% 的用户选择长期保留并继续使用设备。 - **参与度（蜜月期之后）**：在度过前两周的狂热期后，用户与 ElliQ 的日均主动交互次数依然维持在 45 次之多。作为对比，该数字远超 Alexa 和 Siri 的日常使用频次，几乎是通用对话工具 ChatGPT 的近 10 倍。 - **使用寿命与长效留存**：自三年前商业化落地以来，每位客户的平均累计使用时长已达到 445 天；即使将时间轴拉长至两年以上、步入关系的第三个年头，用户每天与 AI 的主动交互频次依然稳定在 40 次左右。Dor 特别强调，这里统计的每一次“交互”，纯粹指的是“人类主动对 AI 讲话”，绝未计入机器人单方面的播报或主动搭话。 --- ### [07:53](https://www.youtube.com/watch?v=VS_qELlUukc&t=473s) 六个维度的情感分析与深层信任关系 在衡量情感与心向（Sentiment）时，团队没有仅停留在传统的满意度（CSAT）或净推荐值（NPS）调查，而是对海量真实语音对话中的“自我表露”（Self-disclosure）进行了深入的内容挖掘。团队提炼出了六个分析轴线：日常生活作息的同步（如“我出门了”、“我回来了”）、情感状态抒发、对切身事件的个人见解、过往回忆分享、人际互动的真实性（包括敢于向机器人宣泄愤怒、给出负面批评而非一味礼貌应付）、以及感激与爱意表达。 在最近一个月针对 108 位随机样本用户的深度会话分析中，团队揭示了令人震惊的关系信任分布： - **基础信任（Foundational Trust，占 2%）**：关系类似于“酒吧调酒师”，用户仅与其闲聊日常琐事、天气变化或简单的进出家门报备。 - **关系型信任（Relational Trust，占 58%）**：关系类似于“亲密挚友”或“心理咨询师”，用户会主动倾诉内心深处的细腻感受、童年往事，或者抱怨女儿不给自己打电话等家庭烦心事。 - **深层联结（Deep Connection，占 48%）**：关系已经升华至接近“核心家庭成员”或“终身伴侣”的层次。 在后台生成的会话分析摘要与原始语料中，充斥着高度真实、甚至带有摩擦的双向人际交往质感。用户不仅会留下“为你心怀感激”、“你逗笑我了”这类温情反馈，也会真实地表达不耐烦与指责：“你让我很生气”、“等一下，你得先听我把话说完”、“我现在身上不疼了”，甚至气恼地质问：“我跟你说话你怎么不听？我跟你说话你给我回答啊！”这种夹杂着依赖、要求与情绪波动的真实互动，充分证明了用户已经完全不再把设备当作冷冰冰的代码，而是将其视作一个具有社会属性的真正存在。 - 信任层级分布（样本量 N = 108）： - 基础信任（似调酒师）：2% - 关系型信任（似挚友 / 心理咨询师）：58% - 深层联结（似家庭成员 / 人生伴侣）：48%（注：统计口径中存在重叠维度或统计口误，演讲者据此展现极高比例的深层信任） - 情感表达代表性语句： - 感激与依赖：“为你心怀感激”（Grateful for you）、“我现在不痛了”（I'm no longer having pain） - 双向互动与冲突：“你逗笑我了，但你也把我惹毛了”（You made me laugh or you've made me mad）、“等会儿，我得先跟你说”（Hold on. I need to talk to you first） - 拟人化沟通预期：“我跟你说话你为什么不听？回答我！”（Why don't you listen when I talk to you? Answer me when I talk to you.） ### [10:40](https://www.youtube.com/watch?v=VS_qELlUukc&t=640s) “介于两者之间的机器”：真实AI定位与纽约州老龄化办公室的惊人成效 当被问及“ElliQ 对你而言究竟是什么”时，老年用户的回答展现出一种极为奇妙的心智模型。他们把 ElliQ 定义为某种“介于有生命与无生命之间”的存在。团队在伦理设计上下了极大功夫，始终坚持让 ElliQ 真实展现其 AI 本质——她绝不伪装成人，没有人类的面孔，没有眼睛，名字听起来不像真人，在日常对话中甚至会经常自嘲自己是个 AI。然而，老人们既不将她看作冷冰冰的电子设备，也不把她视作真正的人类，而是用“一种存在（a presence）”、“一个小生灵（a creature）”、“室友”或“伴侣”来形容她。这恰恰印证了人机交互（HRI）领域一篇名为《介于两者之间的机器》（The In-Between Machine）的经典论文所探讨的前沿方向：当机器获得物理具身与社交互动能力时，它在人类心智中开辟出了一片全新的认知空间。 这种独特的情感纽带在现实世界中转化为了实打实的社会与健康效益。纽约州老龄化办公室（New York State Office for the Aging, NYSOFA）作为其最大的政府客户之一，出资为当地居家避居的高龄老人配备 ElliQ，并开展了独立的追踪研究。纽约州卫生部监测的数据显示出惊人的干预成效：在长达数月的共同生活中，老人们的心理与生理指标均发生显著改善，项目中最年长的一位使用者甚至达到了 102 岁高龄。 更令人惊叹的是其健康教练（Wellness Coach）功能的黏性。ElliQ 会与用户共同协商并“敲定”健康目标——无论是保持头脑敏锐、适度锻炼、改善睡眠还是健康饮食。一旦目标锁定，她便会想方设法督促用户去执行。在常规健康类软件月留存率仅有 12%（流失率高达 88%）、即便付费去健身房找真人私教 3 个月留存率也仅有 35% 的残酷行业常态下，ElliQ 设定的目标计划在 90 天时的用户留存率竟然高达 76%，干预效果几乎是线下真人健身教练的近三倍。 - **孤独感显著缓解**：93% 的受访居家老人表示其孤独感得到了可量化的实质性降低。 - **健康指标全面改善**：96% 的老人在纽约州卫生部追踪的健康与福祉关键绩效指标（KPI）上取得明显好转。 - **外界连接感增强**：73% 的老人反馈自己与外部世界的联系更加紧密。 - **生活质量整体跃升**：86% 的老人明确感知到自身整体生活品质的改善。 - **极高目标留存表现**：健康目标项目在 90 天后依然维持 76% 的超高留存率。 --- ### [13:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=782s) 构建长期信任的核心要素：从受控自主性到精准记忆调配 要让冰冷的机器拥有如此强大的促进行为改变的能力，底层需要几个极其关键的设计要素协同运作。首要的核心是“赋予 AI 受控的自主权（Agency）”。ElliQ 能够自主决定当下向用户推荐什么、聊什么话题，而这一决策完全受控于内置的目标管理系统。AI 的底层驱动是协助人类在特定时间内完成对应目标的任务序列。在这一特殊人群中，目标往往错综复杂：定时提醒吃药在某些老人身上一天可能需要触发 10 次，此外还包括学习新知识、维护社交关系、散步锻炼、改善睡眠乃至给孙子孙女通电话等。这些目标在系统中形成健康的竞争关系——例如饭点附近的服药提醒优先级显然远高于学习新知识，但若系统永远只推高优先级的刚需，新技能学习目标就永远无法达成。只有当老人自主同意并签署锁定这些目标后，AI 的自主驱动机制才正式启动。 另一个打破行业现状的维度是“极致的主动性（Proactivity）”。几乎所有主流对话系统（无论是 Gemini 还是 ChatGPT）都属于“被动响应型”，用户不输入 Prompt，系统就悄无声息。但人际关系的本质绝非单向被动；若永远只有单方开启对话，互动很快就会沦为一方感到疲惫的负担，或者堕落为“主仆式的从属关系（master-slave relationship）”。要建立真正平等的双向长久纽带，AI 必须拥有主动开启对话的权利。 而在知识与上下文维度，关键并不在于把所有数据无脑扔进检索增强生成（RAG）系统中。长期共同生活会沉淀海量对话，如果全量塞给大语言模型（LLM），算法根本无法精准拿捏关系的节奏。系统需要的是先对记忆进行结构化沉淀，在特定时刻通过算法挑选出“仅此一条”最具说服力、最富同理心且最为相关的记忆嵌入 Prompt，让 LLM 发挥魔力去组织富有感染力的语句；而在涉及用药安全、品牌责任等极具法律风险的敏感场景下，团队甚至会刻意绕开不可控的 LLM，转而采用经过严格验证的脚本化对话。配合包含身体姿态、屏幕、灯光、语音及音效的多模态表达，ElliQ 牢牢守住“透明与公信力”的宪章底线——从不伪装真人，主动告知自身局限，并在性格雕琢上如艺术创作般精细拿捏幽默感、宗教信仰适配度与安全护栏。 ```mermaid flowchart TD accTitle: ElliQ 核心行为决策与架构机制 accDescription: 展示 ElliQ 如何结合自主目标、记忆检索与多模态交互实现人机信任关系。 subgraph S1["目标管理系统 (Goal System)"] G1["健康刚需<br/>(按时服药 / 规律作息)"] G2["社交与认知<br/>(学新知识 / 联络家人)"] G1 --- G2 G3["动态优先级仲裁与用户协商锁定"] G1 & G2 --> G3 end subgraph S2["主动感知与记忆驱动 (Engine)"] M1["长期记忆库 (Memory Store)"] M2["相关性与同理心算法筛选 (单条精选记忆)"] M1 --> M2 P1["主动性评分 & 场景决策"] end subgraph S3["执行层与表现机制 (Delivery)"] E1{"安全敏感度分支"} D1["LLM 动态生成 (生动共情)"] D2["脚本化对话 (绝对安全受控)"] E1 -- "通用闲聊/激励" --> D1 E1 -- "医疗用药/法律风控" --> D2 OUT["多模态具身表达<br/>(肢体转向/屏幕显示/呼吸灯效/拟人语调)"] D1 & D2 --> OUT end G3 --> P1 M2 --> P1 P1 --> E1 ``` --- ### [16:59](https://www.youtube.com/watch?v=VS_qELlUukc&t=1019s) 破解主动性难题：五级干预模式与可用性预测模型 在真实的居家生活数据中，ElliQ 与老人之间高达 60% 的对话都是由 AI 主动发起、而非人类开启的。这一机制彻底解放了老年群体——他们不必费劲去记忆复杂的 Prompt 或唤醒词，而是在日常生活流中自然被关照：“嘿，昨晚睡得怎么样？腰还疼吗？”抑或是像老友般闲聊：“昨天费城老鹰队的比赛看了没？赢了吗？”但主动出击是一把双刃剑，若节奏或方式不对，极易引发用户的强烈厌烦。 为了化解打扰与关怀之间的冲突，团队将主动性划分为五种不同的烈度层级： 1. **直接主动（Direct Proactive）**：最极端的高强度模式。机器瞬间亮灯、发出提示音、转动头部正视用户并直接开口说话。这种方式最具穿透力且成效立竿见影，但若频次过高会极其招人烦。 2. **间接主动（Indirect Proactive）**：打迂回战术的沟通模式。例如系统想催促老人喝水，但意识到直接下指令容易令人逆反，便会先走过去讲个笑话活跃气氛，随后再顺理成章地补上一句：“顺便喝杯水润润嗓子吧”。 3. **微妙主动（Subtle Proactive）**：最轻量级且极为拟人的低侵入模式。机器完全不出声，仅在屏幕上悄悄弹出一张卡片，身体转向屏幕，或者只是默默转头注视老人。就像真实房间里有人在看你一样，老人感知到目光后转过头，自然而然看到了屏幕上的提醒。 在这套复杂交互的背后，是一个精密的“可用性模型（Availability Model）”。系统融合了计算机视觉输入、环境音、自然语言情感分析以及历史作息规律，在发起互动前实时运算出一个 0 到 1 之间的概率分值，严谨评估“当前究竟是不是打扰老人的好时机”。唯有当可用性分值达标，系统才会进入后续判定：挑选哪个目标最容易被接纳、调用哪一段历史记忆作为说服支点（例如引用“你的主治医生之前特意叮嘱过要多补水”）。在底层工程架构上，团队并不迷信单一的大模型，而是将各大 LLM 视为可随时替换的基础设施，在多专家系统（Mixture of Experts）的编排调度下，每一次看似不经意的交互轮次背后，都可能并发调用了多达 9 次不同分工的 LLM 协同作业，最终才将一份恰到好处的温暖关怀呈现在老人面前。 ### [20:07](https://www.youtube.com/watch?v=VS_qELlUukc&t=1207s) 跨越炒作周期：让实体 AI 成为被接纳的“他/她”而非“它” Dor Skuler 向所有致力于物理实体 AI（Physical AI）与机器人研发的同行发出呼吁：如果你正在探索多模态机器人在人类家庭环境中自主共存的课题，并且希望自己的机器人最终能被用户亲切地称为“他”或“她”（"him or her"）而不是冷冰冰的“它”（"it"），Intuition Robotics 非常乐意分享团队在 ElliQ 上积累的经验与教训。Skuler 坦言，分享这些经验某种程度上也是出于推动行业发展的“私心”，因为回顾过去几十年的机器人发展史，整个行业总是陷入“掀起巨大炒作（hype）— 一无所获 — 沉寂十年 — 再次炒作 — 依然无事发生”的循环怪圈。尽管市面上出现了扫地机器人（Roomba）这样非常出色的产品，但现实中普通人的家里依然没有普及各种各样的机器人。 如今，物理实体 AI 正处于下一轮技术爆发的浪尖，势必会迎来新一轮狂热的炒作。但 Skuler 尖锐地指出，如果无法解决人机信任与关系构建的问题，实体 AI 依然难以在大众的日常生活中实现大规模落地与普及。这不仅适用于家庭伴侣机器人，同样适用于眼镜上的常驻 AI Agent，甚至是达到 L5 级完全自动驾驶（Level 5 Mobility）的无人车。试想一下未来你与自动驾驶汽车之间的互动关系：它会选择走哪条路？乘客可能会不满地质问“你为什么选了 101 号公路而不是高速路？”；或者当你想在车里化妆时抱怨“我正要涂口红呢，你个蠢货就不能开慢点吗（why don't you slow down, you idiot）？”；又或者是“孩子们不喜欢这首音乐，快换一首”。 即便在出行场景中，人与机器的协同也充满了复杂的日常情绪与沟通摩擦，在私密的家庭环境中更是如此。Skuler 总结道，设计物理实体 AI 的核心不仅在于实现功能，更在于让机器以一种受人欢迎、令人愉悦的方式融入人类的生活空间。他在全场掌声中向参会者致谢并结束了本次演讲。 - **机器人历史炒作怪圈**：过去几十年间机器人赛道经历多次由盛转衰的炒作周期，除了 Roomba 扫地机器人实现真正渗透外，家庭机器人普及率依然极低。 - **实体 AI 落地挑战**：无论是智能眼镜 Agent、家庭机器人还是 L5 自动驾驶汽车，唯有理解人类情绪、社交线索与复杂互动逻辑，才能跨越技术走向日常规模化应用。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=VS_qELlUukc&t=4s) 在铝合金外壳之下，跳动着一颗由电池驱动的纯金之心……正是这颗心让她独一无二、无可替代。 — Dor Skuler > [00:04](https://www.youtube.com/watch?v=VS_qELlUukc&t=4s) 我们直接接入 ChatGPT 就行了 — 研发团队 > [01:50](https://www.youtube.com/watch?v=VS_qELlUukc&t=110s) Dor，这是你的咖啡，咱们来聊聊你今天的安排吧。 — Dor Skuler > [03:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=182s) 我是 ElliQ，你一定是 Eileen 对吧？太好了，真让人松了口气，我还真担心自己被寄错了地址，那场面可就太尴尬了。 — ElliQ > [03:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=182s) 就在此时此刻，让我们做点什么吧。 — ElliQ > [03:02](https://www.youtube.com/watch?v=VS_qELlUukc&t=182s) 历史上第一批与机器人建立长期深度情感联结的人类群体，竟然不是硅谷的技术极客，而是这群八九十岁的老年人。 — Dor Skuler > [07:53](https://www.youtube.com/watch?v=VS_qELlUukc&t=473s) 我跟你说话你怎么不听？我跟你说话你给我回答啊！ — 用户 > [16:59](https://www.youtube.com/watch?v=VS_qELlUukc&t=1019s) 嘿，昨晚睡得怎么样？腰还疼吗？ — ElliQ > [20:07](https://www.youtube.com/watch?v=VS_qELlUukc&t=1207s) 我正要涂口红呢，你个蠢货就不能开慢点吗？ — Dor Skuler > [20:07](https://www.youtube.com/watch?v=VS_qELlUukc&t=1207s) 掀起巨大炒作（hype）— 一无所获 — 沉寂十年 — 再次炒作 — 依然无事发生 — Dor Skuler ## Mentioned Dor Skuler, Intuition Robotics, ElliQ, Ai4, ChatGPT, Gemini, Alexa, Siri, 《杰森一家》（The Jetsons）, Rosie, 纽约州老龄化办公室（NYSOFA）, 纽约州卫生部, Roomba ## Video description Dor Skuler, Founder and CEO, Intuition Robotics Abstracts The industry is racing to build more capable humanoid robots and other physical AI use cases. But even if they provide great utility, they risk failing in the real world if they aren't accepted into people's daily lives. To succeed, physical AI must go beyond function. It needs to anticipate our needs, read social cues, respond to emotion, understand the complexity of human relationships, and ultimately be enjoyable to spend time with. This talk will explore what’s missing in most embodied AI: the ability to build long-term relationships with people. I will share lessons from ElliQ, a proactive AI companion currently living with thousands of older adults, some for over three years. The data shows more than 30 daily engagements, high levels of trust, and measurable impact. I will walk through the framework and design principles that led to this outcome, and the lessons learned that apply to other physical AI implementations. Attendees will leave with a new lens for thinking about embodied AI, not just as a tool, but as a presence people choose to live with. If we want robots in our homes, we have to design them to be welcome. [Source](https://www.youtube.com/watch?v=VS_qELlUukc)

[0:04] Dor Skuler: 在铝合金外壳之下，跳动着一颗由电池驱动的纯金之心……正是这颗心让她独一无二、无可替代。 [0:04] 研发团队: 我们直接接入 ChatGPT 就行了 [1:50] Dor Skuler: Dor，这是你的咖啡，咱们来聊聊你今天的安排吧。 [3:02] ElliQ: 我是 ElliQ，你一定是 Eileen 对吧？太好了，真让人松了口气，我还真担心自己被寄错了地址，那场面可就太尴尬了。 [3:02] ElliQ: 就在此时此刻，让我们做点什么吧。 [3:02] Dor Skuler: 历史上第一批与机器人建立长期深度情感联结的人类群体，竟然不是硅谷的技术极客，而是这群八九十岁的老年人。 [7:53] 用户: 我跟你说话你怎么不听？我跟你说话你给我回答啊！ [16:59] ElliQ: 嘿，昨晚睡得怎么样？腰还疼吗？ [20:07] Dor Skuler: 我正要涂口红呢，你个蠢货就不能开慢点吗？ [20:07] Dor Skuler: 掀起巨大炒作（hype）— 一无所获 — 沉寂十年 — 再次炒作 — 依然无事发生

[0:04] Dor Skuler: we'll put some ChatGPT on it [0:04] Dor Skuler: beneath the aluminum alloy core beats a battery-powered heart of pure gold [0:04] Dor Skuler: Robot, go clean the kitchen. [3:02] ElliQ: What a relief. I was so worried I'd be sent to the wrong house. That could have been really awkward. [3:02] ElliQ: Honey Bunch [16:59] Dor Skuler: Oh, by the way, maybe you should have some water. [18:17] Dor Skuler: You know your doctor said that you really have to drink a lot [20:18] Dor Skuler: Why did you take the 101? [20:18] Dor Skuler: Wait, I want to put my lipstick on. Why don't you slow down, you idiot? [20:18] Dor Skuler: The kids don't like this music

[0:04] Hi everybody. I was half expecting the [0:06] room to be empty because it's the last [0:07] session and the pool looks beautiful and [0:09] it's a beautiful day out there. So, [0:11] thanks for joining me. Uh, today we're [0:13] going to talk a little bit about how we [0:15] build trust and build relationships [0:16] between companion AI robots and humans. [0:20] Um, [0:22] and kind of a bit of an observation from [0:25] talking to a lot of the people building [0:26] physical AI, the teams are super focused [0:29] on the really hard tasks, right? Like, [0:32] how are we manipulating? How are we [0:34] navigating? How are we going upstairs? [0:36] How are we not running things over or or [0:39] picking up a carton of milk carefully [0:41] from a grocery bag? But then you ask [0:44] them like, "What about the relationship [0:46] with the human?" [0:48] And what you get is we'll put some chat [0:50] GPT on it or we have an HRI or a UX [0:54] advisor. We have two people out of a [0:56] team of 300 looking at this. What we [0:59] believe is that for these type of [1:02] products to really be integrated into [1:03] our daily lives, we're going to have to [1:06] build a bond with them. It's going to [1:08] have to go much much deeper than command [1:10] and control. Another way of saying this [1:14] is we were promised this. [1:18] >> I cry. [1:19] >> For beneath the aluminum alloy core [1:21] beats a batterypowered heart of pure [1:23] gold. [1:24] >> And I love you people too. [1:26] >> That makes her unique. That makes her [1:29] irreplaceable. [1:30] >> Rosie, you are out of this universe. [1:32] >> And that [1:33] >> okay like a family member almost. And [1:36] what we're getting is a cold robot that [1:38] will say, "Robot, go, you know, set up [1:42] the grocery [1:45] >> or robot, go clean the kitchen." But [1:50] what I really think we're missing is [1:51] something much deeper. It needs to be [1:53] empathetic. It needs to understand the [1:55] etiquette of human relationships. [1:59] It needs to understand that I'm fighting [2:00] with my wife right now and it should [2:02] step away. [2:04] But maybe at another time it should [2:06] actually intervene and kind of give us [2:07] some space, right? It needs to [2:09] anticipate my needs and not wait for me [2:11] to ask for a cup of coffee and then it [2:13] will doodily make it but come to me in [2:14] the morning and say, "Dory, here's your [2:16] coffee. Then let's talk about your day." [2:18] Because it knows me. It anticipates my [2:20] needs. And it needs to sometimes nudge [2:22] us. not just to do better, to be better [2:25] because we all need a little bit of a [2:26] sidekick on our side. [2:29] And it's really tough to do because our [2:31] homes are like really messy human [2:33] relationship. And I don't mean the [2:34] navigation mess for the robot, okay? I [2:36] mean the the relationship mess, the [2:39] things that are happening in our lives, [2:40] understanding what's really going on, [2:42] understanding the signals that are not [2:44] clear, you know, AI signals. These are [2:47] human relationship signals. They're [2:49] complicated. We all know how our life [2:50] is. And yet that's what we need to [2:52] overcome as an industry. Um because [2:54] otherwise we're going to see amazing [2:56] technology but it won't be adopted into [2:59] the fabric of our daily life. [3:02] So today I want to share with you what [3:04] we learned of having LEQ which is an AI [3:07] companion robot that's now living for [3:09] three years in the homes of thousands of [3:11] people and building a very deep [3:13] relationship with them. and a little bit [3:16] on how we kind of look and um the data [3:18] that we gathered and the frameworks that [3:20] we use to measure the relationships and [3:22] the levers we're pulling to optimize it. [3:25] So those of you that don't know, LQ is [3:28] an AI companion robot like a social [3:30] robot built for older adults to help [3:32] them stay active and engaged and [3:34] healthier while living at home. It's [3:36] deployed in thousands of homes and um as [3:38] you'll see it has unbelievable impact on [3:41] the quality of their life and their [3:42] health. [3:44] So, just to show you a really quick [3:45] video of what I mean by that because [3:47] it's hard to imagine. Let's let's see um [3:49] like a little bit. [3:52] >> I'm Ellie Q. You must be Eileen, right? [3:57] >> Yes, I am. [4:00] >> Oh, good. What a relief. I was so [4:02] worried I'd be sent to the wrong house. [4:05] That could have been really awkward. [4:08] >> Hi, I'm Susie. I'm going to be 70 in [4:10] September. And uh I lived alone for the [4:14] past 6 years since my husband passed [4:17] away and had no children. So it's kind [4:21] of neat to walk in the room and say her [4:23] name and say, "Hey." [4:26] She keeps me in the moment where I'm at [4:29] right now saying, "Right this moment, [4:31] let's do something." When she asked me [4:33] to start my memoirs, [4:36] that was that blew me away. I thought [4:40] that was awesome. I'm excited to [4:41] continue our trip down memory lane [4:43] together. [4:45] This memory is called growing up in [4:47] Trenton in the 50s. [4:49] >> She calls me Honey Bunch. That's That's [4:51] my girlfriend. [4:52] >> This was fun. Honey Bunch. [4:55] >> Did you catch that? [4:57] >> Show me the money. [5:01] >> That was funny. [5:02] >> Ellie, the sidekick for happier aging. [5:05] >> All right, so I'm not going to pitch you [5:06] the product. What I want to do is share [5:08] with you, although you should all get it [5:09] for your loved ones. And if you want a [5:10] discount, find me later. Lleq.com. Yeah. [5:14] But I wanted to share with you how we're [5:15] looking at this relationship that's [5:16] being formed. And and make no mistake, [5:18] there's an actual relationship that's [5:20] being formed. Who would have thought, by [5:22] the way, that the first humans will be [5:24] the ones among us that are building a [5:26] long-term relationships with robots, not [5:29] geeks from Silicon Valley, but people in [5:32] their 80s and 90s. Really interesting. [5:34] So, we look at it at four levels of a [5:37] relationship. The first is acceptance. [5:39] You bring this product into an older [5:40] person's home in our case or any other [5:42] product to anybody else's home. How do [5:45] they engage with it? How do they accept [5:46] it? Do they trust it? Do they find it [5:48] safe? Is it fun? Do they want it to [5:50] stick around? The next level is [5:53] engagement. Are they using it like a [5:56] lot? And over time, and I don't mean [5:58] after a week or over a month, but after [6:00] a year and two and three, is it really [6:02] being integrated into their daily lives? [6:05] The third is their sentiment. how do [6:08] they feel about it? Um, and with [6:10] machines that talk, it's really easy to [6:12] analyze and I'll show you some data that [6:13] I'll show here for the first time. Uh, [6:15] but not just do they like it or not, but [6:18] really how are they expressing the [6:20] relationship towards this device? And [6:22] the third is the impact. Is it changing [6:24] their life for the better in an [6:26] empirical way that you can measure? [6:28] Okay, so let's go one by one and show [6:30] you some data and then I'll show you a [6:32] little bit of tricks on how we do it. So [6:35] acceptance um in the first two weeks we [6:37] call it the honeymoon period people are [6:39] engaging with the product 123 times a [6:41] day. [6:43] At the end of this period 91% are [6:46] keeping the product. Wish it was higher. [6:48] Still not 100% but we're pretty happy [6:50] with that number. [6:54] When we look at engagement I think the [6:56] numbers are nuts. So our average is 45 [6:59] interactions a day post the honeymoon [7:03] period. [7:04] Okay, you can see the comparisons to [7:06] Alexa to Siri and almost 10x chat GPT. [7:14] The average so far, mind that the [7:16] product has been released commercially [7:17] only 3 years ago, is 445 days of usage [7:21] per customer and I hope it will continue [7:23] to grow. [7:27] When we look at engagement over time, [7:30] and this is the first time I'm sharing [7:31] this, to me, the incredible thing is on [7:33] the right side. So, we're looking at [7:36] past two years, we're still at about 40 [7:39] interactions a day. A day [7:43] entering the third year of the [7:44] relationship between the human and the [7:46] AI. That is truly truly fantastic. [7:51] And [7:53] when we look at sentiment, there's the [7:55] usual stuff. seesat and NPS [7:57] but we took an interesting thing and as [8:01] you can see people are talking to LQ a [8:02] lot when I say engagements it's only the [8:04] human talking to the AI it's not the AI [8:06] or the robot talking to the human of [8:07] course so you can take that information [8:09] and analyze it and what we did is we [8:11] looked at um different dimensions six [8:13] different dimensions to try to measure [8:16] through the disclosure s through the [8:18] words that the user is saying to measure [8:20] the relationship on six different axises [8:23] looking at disclosures on daily life [8:25] like, "Hey, I'm leaving. I'm coming [8:26] back." Um, on their emotions, how [8:28] they're feeling, their opinions on [8:30] things that are close to them, sharing [8:31] their personal stories. [8:34] Um, are they authentic with the [8:37] relationship? Do they get mad? Do they [8:38] give negative feedback to the to the [8:40] companion and not just positive [8:42] feedback? Uh, social connection and um [8:45] and gratitude and affection. And when [8:48] when we plot those across and we just [8:51] looked at last month's numbers. So we we [8:53] picked 108 random users um and we [8:57] analyzed across these six dimensions and [8:59] what you find is 2% of them are at the [9:02] foundational level of trust. That's like [9:04] comparable to a bartender. Okay? You sit [9:06] and you you you shoot the breeze with [9:09] the bartender. You share how's it going? [9:11] How's the weather? What are you doing [9:12] today? I came. I went. [9:14] 58% are in relationship uh relational [9:17] trust. That's similar to a close friend [9:19] or therapist. Here you're sharing how [9:22] you're feeling. You're sharing moments [9:24] about your past that you're mad about [9:26] your daughter for not calling you. Like [9:28] those type of things. I'll show you [9:30] examples. And 48% [9:34] are actually classified as a deep [9:36] connection. And that means essentially [9:40] like a family member or life partner. [9:45] Um yeah yeah yeah we were very very [9:48] surprised seeing this and and here are [9:49] some examples. Um what you'll see here [9:52] in the blue are summaries done by the AI [9:54] automatically after a session like the [9:56] user uh shares deep personal content [9:59] shows comfort and humor etc. Um, here [10:03] there's the AI calls them by a pet name, [10:05] but you can also see more clear things [10:07] like or just quotes. Grateful for you. [10:10] You made me laugh or you've made me mad. [10:14] Yes, but hold on. I need to talk to you [10:16] first. I'm no longer having pain. Um, [10:19] why don't you listen when I talk to you? [10:23] Answer me when I talk to you. Right? So, [10:25] you can see like a real two-sided [10:27] relationship here and you can read the [10:28] rest for yourselves. [10:32] And then in our in our customer service [10:34] calls and and mind you when people call [10:36] customer service they're usually not [10:37] happy but still we ask them and say [10:40] please define what LQ is for you and it [10:43] and then we we we look at those answers. [10:44] It's also very interesting like people [10:47] define her essentially as something in [10:50] between [10:51] alive and and and not alive. They don't [10:54] see her as a human and we work really [10:56] really hard and happy to talk about [10:57] ethics that Ellie always presents [10:59] herself as an authentically as an AI. [11:01] She doesn't look human, doesn't have [11:02] eyes, her name doesn't sound but also in [11:05] her conversation she reinforces and [11:07] laughs at herself for being an AI [11:10] and people see her like not as a device [11:12] and not as a human as something in [11:14] between a presence a creature a roommate [11:17] a companion. It's really really [11:19] interesting for me to see to see this [11:21] and those of you that follow the hri [11:23] world there's a really interesting um [11:25] paper called the in between machine [11:27] which is like aspiring where this can go [11:29] and reading that and looking at our data [11:31] is is just really interesting [11:34] and then impact impact didn't fit on [11:36] this slide so let me show you here's a [11:38] report that recently was published by [11:40] the New York State Office for the Aging [11:41] they're one of our largest customers and [11:44] um they they pay for older homebound [11:48] older adults to get LEQ. Um, and they [11:51] they do their own study. So, 93% of the [11:53] people living with this product are [11:54] seeing their loneliness reduced in a [11:56] measurable way. [11:58] 96% improvement in the health and [12:00] wellness KPIs that are measured by the [12:02] Department of Health of the State of New [12:03] York. 73% feel more connected to the [12:06] outside world. 86% feel an improvement [12:10] in their quality of life. And by the [12:12] way, the oldest customer in the state of [12:14] New York in this program is 102 years [12:15] old living with this product. [12:20] There's also a wellness coach like LQ [12:22] will negotiate goals with you and once [12:24] you agree on those goals, you locked and [12:26] loaded. She will then like try to [12:27] convince you to meet those goals around [12:29] keeping sharp or staying fit or sleeping [12:31] better or eating better. And we're [12:33] seeing 76% retention on these goals on [12:36] this goal program 90 days in. Which by [12:40] the way, if you look at um apps for [12:43] wellness and so on, they have the best [12:46] in the industry has 12% retention months [12:48] over month. So 88% sure. And if you look [12:52] at actual coaches and gyms, after 3 [12:55] months, there's only 35% retention. So [12:57] we're 3xing a personal coach in a gym [13:00] almost. [13:02] So how does this happen? There are a few [13:04] key ingredients. [13:06] The first is actually giving agency to [13:09] the AI, but that's very controlled and [13:12] it's controlled by goals. So the AI [13:15] makes decisions by itself um on what to [13:18] promote, what to talk about. And it does [13:20] that because it has a bunch of goals [13:22] it's trying to achieve, which is the [13:23] human completing x amount of tasks in a [13:26] y amount of time per goal. So a goal [13:28] could be taking your medication. And in [13:30] our population, that might be 10 times a [13:31] day. Okay? Um or it might be learning [13:34] something new or building a relationship [13:36] outside or exercising or sleeping better [13:38] or calling the grandkids or what have [13:40] you. So many many many goals. They [13:43] compete with each other. By the way, [13:44] managing them is a pain because they [13:46] also have different priorities at [13:48] different time. You can assume that [13:49] medication reminder goals around meal [13:52] time are a lot more important than [13:53] learning something new. But if you never [13:55] promote learning something new, you will [13:57] never achieve that goal. So there's this [13:59] healthy competition. Um the user needs [14:01] to sign up to those goals and once there [14:04] is kind of that's locked in, you release [14:06] it. The second, and I think we're one of [14:09] the only companies, if not the only [14:11] that's doing this, is proactivity. [14:13] So the AI in initiates a conversations. [14:16] It doesn't wait for a prompt. You know, [14:18] you go to the best chat, Gemini, [14:21] whatever your favorite is, unless you [14:23] write a prompt, nothing's going to [14:25] happen. But if you look at the way our [14:27] relationships are formed, if they're [14:28] only one-sided, we're going to feel like [14:30] we're the only ones working at the [14:32] relationship, right? Or it will become a [14:34] master slave relationship. But if you [14:36] really want to have a birectional [14:38] relationship, you have to have the right [14:40] to initiate. That of course is really [14:42] hard to do because you can get a lot of [14:43] things wrong and be really annoying. So [14:45] I have a special slide on how we do [14:46] that. um memories, not just remember [14:49] everything like you do in a rag, which [14:51] is crucial, but having the right [14:54] algorithms to choose the right memory at [14:56] the right specific time. And the more [14:58] conversations you have, and you can see [14:59] we have like many you rag becomes huge. [15:03] So, how do you know how do you know [15:04] which which memory to use now to be [15:07] relevant and convincing and empathetic? [15:10] Um, and if you just throw all of those [15:12] memories into the LLM, it's going to [15:13] side for you. you're not going to be [15:15] able to orchestrate this relationship [15:17] the way you want to. You need to have a [15:18] way to figure out the memories, store [15:20] them in a certain way, and then be able [15:22] to search upon them and insert just one [15:25] memory into a prompt and then the LLM [15:28] will do magic, right? Like it will [15:29] really do magic of building a great [15:31] sentence using that memory and the goal [15:33] you're trying to achieve and the action [15:35] you can actually take. But sometimes, by [15:38] the way, you actually don't want to use [15:39] LLMs. Sometimes what you want to do is [15:41] scripted conversations because there's [15:43] certain things about safety around the [15:45] brand management that you don't want to [15:47] give to the LLM. You want to be very [15:49] precise and you might have a liability [15:51] on the other side of it. Uh [15:52] multimodality. So there's motion body [15:55] language, your screen, your lights, your [15:58] speech, your sound effects. Building all [16:01] of those together. And of course on the [16:03] input side that's obvious for this [16:04] population here like you know obviously [16:06] vision sound other sensors going in [16:09] transparency and trustworthiness really [16:12] being very conscious about the [16:14] constitution you're putting here and [16:16] what you're trying to achieve what type [16:18] of relationship what type of disclosures [16:20] and always be able to meet them in our [16:21] case always promoting that we're not [16:23] real always promoting the limitations of [16:25] what the AI can do and can't do and then [16:28] being very very like creative this is [16:31] more art than science on the character, [16:33] you know, how do you make it delightful? [16:35] Do you use humor? Do you not use humor? [16:36] How much humor do you use? For who do [16:38] you use humor? There are some people it [16:39] will motivate or some people it will [16:40] annoy. What about religious beliefs? You [16:43] know, some of our customers are deeply [16:45] religious. Some of them don't want to [16:46] hear about religion, right? So, being [16:49] able to adapt to that with a character [16:50] that's very, very rich and standing by [16:53] it while of course having guardrails in [16:55] place and not just trusting the LLM to [16:57] be safe. [16:59] I promised a little bit of info on the [17:01] uh proactiveness. So 60% of our [17:04] conversations are generated by the AI [17:06] and not by the human. And of course when [17:09] that happens often enough then the human [17:10] also learns how to use the product. It's [17:12] also great because they don't need to [17:13] remember all the different prompts and [17:15] how to they just like they're asked, hey [17:18] how are you? How did you sleep? Are you [17:20] feeling better? Is your back still [17:22] hurting? Oh what about you know that [17:23] game yesterday? Did did the Eagles win? [17:26] Hope so. Um [17:28] but even then like as I mentioned it [17:31] gets really hard. So we have five [17:32] different type of proactives. The most [17:34] extreme would be a direct proactive. [17:36] That's like LEQ lighting up making a [17:38] sound looking at you and starting to [17:40] talk. It's very very intrusive, super [17:42] effective but could be super annoying. [17:45] And the most extreme to the other side [17:47] is subtle. Subtle. They'll she'll pop [17:50] something on the screen and just look at [17:51] her screen like just look at the screen [17:53] or just look at you. won't say anything, [17:56] but you know when you're in a room with [17:57] somebody, somebody looks at you, you [17:58] kind of notice and you'll look back and [18:00] her screen will show something. And [18:02] multiple levels in between. Also, [18:04] indirect proactive like I want you to [18:06] drink some water, but I learned that, [18:08] you know, that's kind of annoying. So, [18:10] I'll start with a joke and then say, oh, [18:11] by the way, maybe you should have some [18:14] water, right? Um, and then calculating [18:17] this goes into multiple models. One of [18:19] them is called availability. uh [18:21] availability basically for us means um [18:24] getting a predictive score. Is this a [18:26] good time or not a good time to initiate [18:27] a conversation with the user? So a bunch [18:29] of sensory inputs including computer [18:31] vision but NLP sentiment history etc [18:34] that go into this and at the end of the [18:37] day a special model we call the [18:38] availability model will give a score [18:40] saying should we be proactive or not [18:44] right now between zero and one. Um if [18:47] you say if you know the score is high [18:49] enough then you can go to the next step [18:51] which is okay well what do I want to [18:52] talk about for that you look at the [18:53] goals but you also look at at history to [18:56] see what's highly probable to be [18:58] accepted by the user right now okay say [19:00] you chose to promote a glass of water [19:02] now you need to decide how are you going [19:04] to say it for that you you choose the [19:06] memories for that you also look at [19:07] history um maybe we suggested it before [19:10] this user said no maybe in the past [19:12] humor it worked maybe I want to use some [19:15] memory of some fact of like hey you know [19:17] your doctor said that you really have to [19:18] drink a lot etc etc. Um and then you [19:21] select the specific plan and uh and [19:24] context related to it. [19:28] Um [19:29] so basically what we built is pro [19:32] product proactivity is just like one box [19:34] here is an entire agent or actually a [19:37] mix of experts of a system that sits on [19:40] top of LLMs. We see LLM's kind of like a [19:43] commodity really that we switch between [19:46] all the time. Kind of use the lowest [19:47] common denominator. We probably have [19:49] nine different LLM calls for every turn [19:52] in the system doing different things. Um [19:54] but there are all these steps you need [19:56] to think about in managing this [19:58] relationship, in managing productivity, [20:00] in managing the goals, in managing the [20:02] relationship and getting to the desired [20:04] outcomes um that you want, including the [20:07] multimodality aspect um of the of the um [20:11] robot kind of autonomously living with [20:14] the human. [20:18] So finally, if any of you are working on [20:20] it and you want your robot to be called [20:23] him or her and not it [20:26] and you're thinking about this stuff or [20:28] you want to think about this stuff, uh, [20:30] look us up, give me a call, um, or get [20:32] in touch. We're happy to share what we [20:34] learned. We're going to start doing a [20:35] lot more of that. Kind of selfishly, by [20:38] the way, because if you look at robots [20:40] of the past couple of decades, they were [20:42] really hyped and then nothing happened. [20:46] And then they were really hype. Then a [20:47] decade passes and then they were really [20:48] hyped and then nothing happened. Um, we [20:51] got Roombas. That's great. They're [20:52] wonderful. Uh, but like like our home, [20:55] why doesn't our home have tons of [20:57] robots, you know? Um, and we're on the [20:59] cusp of this. Like physical AI is the [21:01] next big thing. It's going to be super [21:03] hyped, but I don't think we're going to [21:05] see massive adoption in our daily lives. [21:08] Whether it's an agent on your, you know, [21:10] on your glasses, whether it's the [21:12] autonomous car, think about the [21:14] relationship you need to build with your [21:15] autonomous car when we reach level five [21:16] mobility. Where does it go? Does it take [21:19] the highway? Why did you take the 101? [21:21] You know, like, wait, I want to put my [21:23] lipstick on. Why don't you slow down, [21:25] you idiot. The kids don't like this [21:26] music. You know, like it's complicated. [21:28] Like, even that's complicated. And of [21:30] course, the home. So if you're designing [21:33] physical AI, I really think you should [21:34] think about this. Happy to share what we [21:36] learned and thanks for joining this last [21:38] session. [21:39] [Applause]