Podcast Atlas
concept Updated 2026-07-09 Tags: Robotics, Physical-Ai, Investment

Embodied AI

Embodied AI refers to physical AI, robotics, and systems that act in real-world environments. 高手怎么用 AI？普通人怎么学 AI？投资人如何投 AI？｜对谈课代表立正 treats the area as highly valued and bubbly, but also important enough that the host does not want to dismiss every participant as speculative. 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】 adds a concrete consumer case through Xiaoban, where the embodied AI problem is household companionship rather than industrial work or robot performance.

从会跳舞到有感知，触觉是机器人通往智能的门票吗？｜ S10E19 adds a tactile-infrastructure view through Yimu Technology / 一目科技. Eric Li Zhiqiang / 李志强 argues that robots already look more mature in mobility than in dexterous hands or physical understanding, because vision cannot supply fine contact precision, occlusion-resistant feedback, or force control. The source therefore adds Tactile Sensing, Optical Tactile Sensing, TouchNet, and Tactile Transformer Encoder as a physical access layer between robot bodies and model brains.

具身智能的滔天大泡沫中，他已经把机器人送进300个家庭｜对话张翼：未来不远创始人/CEO adds a task-and-service household case through Weilai Buyuan and F2 Home Robot. Zhang Yi argues that Home Service Robots have to work inside real families, earn repeated use, collect Household Robot Data Flywheel signals, and reduce cost through Consumer Robotics Full Stack tradeoffs such as wheel bases, two-claw hands, and home-appropriate component specs.

哪条路线，才能通往「世界模型」的终局？｜对话黄碧薇：Aether AI 创始人 adds an embodied-intelligence foundation-model view through Aether AI. Huang Biwei argues that robotics remains early because physical action requires generalization under hidden variables, data bias, missing values, and distribution shift. Her proposed route is Causal World Models trained from simulated data, egocentric data, video data, and teleoperation data.

131. 印奇出任阶跃星辰董事长的访谈：聪明人的诱惑、取舍、超长链路残酷淘汰赛、阶跃函数和超多元方程 adds Yin Qi’s commercialization and AGI view. He argues that AGI must interact with the physical world, but that the route may pass through cars, phones, wearables, and other terminals before full robots. In that view, Qianli Technology is the first vehicle terminal and StepFun supplies the foundation-model “brain” through AI Plus Terminals.

132. 对星海图创始人高继扬的3小时访谈：鲶鱼、曾国藩、Waymo与Momenta的两面、一只狼与许华哲的离开 adds an industrial/productivity robotics case through Xinghaitu. Gao Jiyang argues that the robot body is both the data carrier and the commercial good, so embodied intelligence has to combine Physical World Data Flywheel, Real Robot Data Strategy, Embodied AI Value Chain, Wheel-Based Dual-Arm Robots, and Production Robot Scenario Selection rather than only a detached robot brain.

166: 许华哲再次具身创业：不想错过最大的西瓜 adds Xu Huazhe’s founder route through Poke Robotics. He agrees that bodies and data matter, but argues that the key prize is Physical AGI: a general robot brain that can act across household tasks. This creates a useful contrast with industrial-first embodied-AI paths, because Xu worries that shipment volume, production landing, data sales, or demo performance can distract companies from AI Native Robotics and Unified Robot Models.

133. 对谢赛宁的7小时马拉松访谈：世界模型、逃出硅谷、AMI Labs、两次拒绝Ilya、杨立昆、李飞飞和42 adds a broader physical-world intelligence bridge through AMI Labs. Xie Saining does not reduce the problem to robots, but argues that World Models need continuous perception, action-conditioned prediction, memory, planning, and real-world partner data from domains such as wearables, hospitals, farms, cars, and other physical contexts.

134. 【数据的综述】和谢晨聊，新时代的石油、历史、版图、数据金字塔、定价与Recipe adds 谢晨 and 光轮智能 as the data-infrastructure view of embodied AI. Xie argues that real robot data remains necessary but cannot scale alone, so the field needs Embodied Data Pyramid, Robotics Simulation Evaluation, Data Engine Learning Loop, and Data Recipe Co-Creation to provide training, evaluation, and feedback at industrial scale.

143. 对何小鹏的第二次访谈：更大赌注、人形机器人Iron诞生、那场意外、技术剧变下CEO、GX和缝合怪 adds He Xiaopeng / 何小鹏 and XPeng / 小鹏汽车 as the car-plus-humanoid version. The source names this broader frame Physical AI, where intelligent vehicles, XPeng Iron, XPeng GX, data, compute, hardware, controls, manufacturing, and AI Organization Design are treated as one system. Its main caution is Stitched AI Architecture: adding rules and partial AI to an old stack can raise performance without reaching generalized physical intelligence.

144. 对杨萌的4小时访谈：消费电子死与生、第三类公司、端侧模型、产品方法、游戏模式 adds Anker Innovations / 安克创新 as a consumer-electronics route into embodied intelligence. Yang Meng / 杨萌 places current work in staged robotics: flat-ground robots such as cleaners and mowers, then three-dimensional interaction robots such as a security “watchdog,” and only later humanoids when technical route, leader, and organization are ready.

170: 【具身季报 26Q2】世界模型大风不停，和不想被贴标签的人 adds a quarterly industry-map view through LateTalk and Chen Zhe Peter. The source ties Humanoid Robot Marathon, Robot Logistics Sorting, Dexterous Manipulation, Embodied Robot Data Paradigms, and World Model VLA Fusion into one embodied-AI competition: reliable robot bodies, affordable hands, real industrial scenes, model route, data collection, and foundation-lab entry all shape who controls the robot brain and business layer.

Commercialization Mentioned

  • Robot performances and rentals.
  • Data collection.
  • Some industrial scenarios.
  • Sales to research institutions.
  • Consumer Companion Robots such as Xiaoban, where value depends on Robot Liveliness, safety, materials, motion, memory, and emotional interaction.
  • Home Service Robots such as F2 Home Robot, where value depends on child care support, light chores, household safety, maintenance intervals, and real family data.
  • Generalized robot brains built from Causal World Models, where product feasibility depends on data, compute, and robotics full-stack engineering.
  • Cars, cabins, phones, and wearables as staged terminals for physical-world interaction before mature robots.
  • Developer and production robots such as Xinghaitu’s Wheel-Based Dual-Arm Robots, where value depends on real-world data, whole-machine execution, model/action architecture, and production-scene fit.
  • General household robots such as Poke Robotics’ proposed first product, where value depends on Physical AGI, Robot Active Use Metrics, product safety boundaries, and model generalization rather than only a single task wedge.
  • World-model partner environments such as wearables, hospitals, farms, cars, and other physical contexts where AMI Labs expects perception, action, and feedback loops to matter.
  • Simulation, evaluation, and data-engine companies such as 光轮智能, where value depends on physically meaningful environments, recipe discovery, and model-team co-iteration.
  • Vehicle and humanoid systems such as XPeng GX and XPeng Iron, where value depends on full-stack physical AI, lower-bound safety, motion control, manufacturing reliability, and public trust.
  • Household security robots such as Anker’s proposed Household Security Robots, where value depends on closing the loop from detection to response in a trusted home-security context.
  • Logistics sorting, dexterous manipulation, and humanoid stress-test events where companies such as Figure AI, Xingdong Era, Honor, 5G Robotics, and Genesis Robotics try to turn demos into data and customer proof.
  • Tactile sensing for robot hands, industrial insertion/assembly, screwdriving, and medical manipulation, where value depends on force feedback, slip detection, texture, and real-time correction rather than only visual recognition.

Connections