超维动力押注人形机器人全栈：KAI Bot、灵巧手与数据闭环一起推进 X Square Robot Bets on a Full-Stack Path for Humanoid Robots
这家成立于2025年的深圳公司把机器人本体、灵巧手、数据采集硬件、世界模型和训练平台放在同一条链路中，重点不只是参数，而是跨场景学习、跨本体适配和商业化入口。 The Shenzhen startup, founded in 2025, is linking humanoid hardware, dexterous hands, data-capture devices, world models and training infrastructure into one loop, with the larger test being reliability, deployment and reusable robot skills.
人形机器人行业的关注点正在从单个炫技动作转向可复现能力：换一个环境、换一台机器人，系统还能不能稳定完成同一任务。量子位报道称，深圳创业公司超维动力选择从创立初期就做全栈布局，试图把本体、模型、真实数据和应用场景连成闭环。
其产品线包括全尺寸人形机器人KAI Bot、灵巧手KAI Hand、数据采集设备KAI Halo、KAI World Model以及训练平台KAI Embodied AI Infra。这里的“全栈”并非单纯扩展业务范围，而是围绕具身智能在真实世界中反复学习、验证和部署的需求搭建系统。
硬件参数服务于示范学习
KAI Bot身高173厘米、体重70公斤，全身117个自由度，近八成身体覆盖触觉皮肤，约有1.8万个触觉触点，理论上可感知0.1牛顿以上的轻微触碰。KAI Hand则拥有37个自由度，指尖力超过30牛顿，连续抓握10分钟后整手最高温度不超过人体体温。
这些设计的重点不是让机器人外形更像人，而是让它能更完整地复现人类示范中的肩、腰、手等动作。对AI工具和机器人开发者来说，这意味着硬件自由度、触觉反馈和热管理会直接影响数据质量与可训练动作范围。
SMASH展示从感知到行动的链路
超维动力的乒乓球系统SMASH用于展示模型如何把视觉感知转成实时动作。系统可在毫秒级捕捉来球、预测轨迹，并结合视觉信息与机器人自身状态计算击球时机、速度和角度。报道提到，SMASH已经完成机器人之间的11分制完整对局。
更值得关注的是，相关算法还适配了宇树G1和智元远征A3。跨本体适配说明这套能力并不只绑定自家机器人，也让它对具身智能工具链、模型评测和机器人导航站用户更有参考价值。
数据与训练平台构成闭环入口
KAI Halo让数百名采集人员进入家庭、超市、商业空间和小型工厂，记录第一人称多模态数据。公司披露，目前已有超过10万小时视频，覆盖20多个场景和300多种全身原子技能。
KAI Embodied AI Infra负责数据处理、模型训练、仿真评测和真机部署，并把机器人执行结果回传到下一轮训练。按官方数据，高质量数据生产效率可提升10倍，模型训练与评测速度平均提升5至7倍。
商业化上，公司没有等待完整系统成熟后再推出产品。KAI Hand和KAI Halo Lite已在2026年WAIC期间公开销售，灵巧手可服务其他机器人，采集设备和Infra也可作为对外业务入口，同时反哺更多场景与数据。
风险仍然明确：这是一家成立约一年的公司，仍需面对长时间运行可靠性、批量交付、资金和组织压力。最终决定其竞争力的，不是某个参数或一场演示，而是整套系统能否持续稳定运转并创造实际价值。
The humanoid robotics market is shifting from flashy one-off demonstrations to reproducible capability: can the same task still work when the environment or robot body changes? According to QbitAI, Shenzhen startup X Square Robot is taking a full-stack route from an early stage, linking robot hardware, models, real-world data and applications into one loop.
Its product lineup includes the full-size KAI Bot humanoid, KAI Hand dexterous hand, KAI Halo data-capture device, KAI World Model and KAI Embodied AI Infra training platform. In this context, full stack is less about broad product sprawl and more about supporting repeated learning, validation and deployment in the physical world.
Hardware Built for Demonstration Learning
KAI Bot is 173 cm tall, weighs 70 kg, has 117 degrees of freedom and covers nearly 80% of its body with tactile skin. It has about 18,000 tactile points and is designed to detect light touches above 0.1 newton. KAI Hand has 37 degrees of freedom, fingertip force above 30 newtons and, after 10 minutes of continuous grasping, a maximum whole-hand temperature no higher than human body temperature.
The point is not simply to look more human. The design is intended to help the robot reproduce human demonstrations across shoulders, waist, hands and other joints. For AI tool and robotics developers, that makes degrees of freedom, tactile feedback and thermal control directly relevant to data quality and the range of trainable actions.
SMASH Links Perception to Action
X Square Robot’s table-tennis system SMASH shows how a model can move from visual perception to real-time action. It can capture an incoming ball at millisecond level, predict its trajectory, then combine visual input with the robot’s own state to calculate timing, speed and angle for a shot. The report says SMASH has completed full 11-point matches between robots.
A notable detail is that the related algorithms have also been adapted to Unitree G1 and Agibot Expedition A3. This cross-embodiment support suggests the capability is not limited to the company’s own robot, which makes it more relevant for embodied AI tooling, model evaluation and robot resource directories.
Data and Infrastructure Form the Loop
KAI Halo sends hundreds of data collectors into homes, supermarkets, commercial spaces and small factories to record first-person multimodal data. The company says it has more than 100,000 hours of video across over 20 scenarios and more than 300 whole-body atomic skills.
KAI Embodied AI Infra handles data processing, model training, simulation evaluation and real-machine deployment, then feeds execution results back into the next training cycle. Based on official figures, high-quality data production efficiency can improve by 10x, while model training and evaluation speed can improve by an average of 5x to 7x.
The company is also commercializing before the full system is mature. KAI Hand and KAI Halo Lite were publicly sold during WAIC 2026. The dexterous hand can serve other robots, while the capture device and Infra platform can become external business entry points and bring back more scenarios and data.
The risks remain concrete. As a company founded around a year ago, X Square Robot still has to prove long-duration reliability, batch delivery, funding resilience and organizational execution. Its future will depend less on any single specification or demo and more on whether the full system can run reliably and create sustained value.
来源
- 量子位 · 08-26 11:02