36Kr Exclusive | Zhejiang University Professor Team Secures Investment from Cailong and SenseTime, Developing AI Brain for Robots in High-Risk Scenarios 36氪首发 | 浙大教授团队获财通、商汤投资，做高危场景具身机器人大脑

While many startups are still packaging their robots under the banner of "intelligent inspection," Shu Jiangpeng of Hangzhou Kuangxing Technology has bluntly stated the reality: most current robots remain stuck in the "seeing" stage—they can detect problems but cannot solve them, essentially functioning as advanced cameras. This critique sharply highlights the awkward position of embodied intelligence in industrial applications: we've built so many mobile devices, but do they truly "understand" engineering?

Kuangxing Technology has recently secured tens of millions of yuan in Pre-A round financing, with investors including Caitong Capital and SenseTime’s Guoxiang Capital. While the amount isn't huge, the direction it signals is noteworthy. Instead of manufacturing robot hardware, the company focuses on developing the "brain"—an engineering multimodal large model capable of integrating data from images, point clouds, ultrasound, and other sources to perform sub-millimeter quantitative identification of defects like concrete cracks and steel corrosion. In other words, it aims to equip robots with an "engineer's" mindset rather than just an "inspector's" mobility. Shu Jiangpeng points out directly that the industry's core pain point is "strong inspection, weak diagnosis"—robots spot cracks but still rely on human verification, which he considers an insult to AI.

A key technological advantage lies in data barriers. Kuangxing claims to possess the world's largest structured multimodal database in the infrastructure sector, containing millions of professionally annotated negative samples—two orders of magnitude more than competitors. This may sound impressive, but it stems from 15 years of accumulated research at Zhejiang University's civil engineering department. In AI, data is oil, and domain-specific professional data is rare fuel. Competitors relying solely on public datasets for training will struggle to match its diagnostic precision. However, a subtlety lies beneath: collecting negative sample data in industrial scenarios is inherently difficult. Kuangxing, leveraging its academic background, has tapped into this goldmine, while other teams may need to start from scratch to catch up.

Regarding commercialization, Kuangxing has chosen an asset-light model—purchasing or collaborating with manufacturers for robot hardware and adding its self-developed "brain box." This is a smart move, avoiding the heavy investment traps of hardware manufacturing while focusing on software and algorithm output. But the risk is clear: if hardware suppliers develop their own AI capabilities, how wide will Kuangxing's competitive moat remain? Currently, it has signed contracts worth tens of millions of yuan with leading enterprises like Jiangxi Copper and State Grid, with underground four-legged robots already in pilot operations. Shu Jiangpeng’s economic rationale is pragmatic: a robot costing a few hundred thousand yuan can replace human labor costing over a million annually while avoiding production stoppage risks. This logic holds, but only if the robots are sufficiently reliable. He admits that current coverage is only 50%, with the other half still relying on manual work. Moving from 50% to 90% coverage isn’t something that can be achieved simply by refining algorithms—underground network fluctuations and narrow tunnel adjustments in mines can easily cause robots to malfunction. The battleground for embodied intelligence has never been in laboratories but in messy real-world conditions.

On the debate surrounding humanoid robots, Shu Jiangpeng takes a pragmatic stance: short-term efforts will focus on four-legged robots due to their greater stability, as losses from falls in high-risk scenarios are unacceptable. This casts a sobering note on the current rush to develop humanoid robots. What industry needs isn’t human-shaped showpieces but robust tools capable of tackling tough tasks. Humanoid robots might find their place eventually, but expecting them to conduct inspections in mines within three to five years is unrealistic. Kuangxing has made it clear that building robot hardware isn’t its priority—it focuses solely on the "brain." This specialization is commendable: after all, expertise matters, and AI companies trying to compete in hardware often meet unfortunate ends.

Investors’ perspectives also offer food for thought. SenseTime’s Guoxiang Capital is bullish on an "air-ground-mine integrated" system solution, while Caitong Capital emphasizes "hand-brain-eye coordination." Yet beneath these terms, the core bet is that embodied intelligence can carve out a niche in traditional industries. The trend of using special robots to replace human labor is irreversible, especially today when young people prefer delivering food to working in mines. However, investment booms often bring bubbles: many companies secure funding with polished presentations but fail to deliver even one stable operational case. Kuangxing’s ability to reach contract implementation at least demonstrates that its technology isn’t just theoretical.

Overall, Kuangxing Technology’s approach is on the right track: industrial robots shouldn’t just serve as mobile sensors—they need diagnostic and intervention capabilities. Yet the road ahead remains thorny. Data accumulation takes time, algorithm generalization faces scenario-specific challenges, and commercialization still has to overcome barriers of cost and trust. Shu Jiangpeng’s goal to "build the robot brain that best understands engineering" is inspiring, but the complexity of the engineering world far exceeds the logic of code. Perhaps this breakthrough battle has just begun, and we need more companies like Kuangxing—using solid data and precise pain-point solutions to pull AI out of illusions and back into reality.

当一堆创业公司还在用“智能巡检”包装自己的机器人时，杭州旷行科技的舒江鹏直接把话挑明了：现在的机器人大多停留在“看”的阶段，能发现问题却解决不了问题，本质上还是个高级摄像头。这番吐槽精准戳中了具身智能在工业落地中的尴尬——我们造了这么多会移动的设备，但它们真的“懂”工程吗？

旷行科技刚拿到数千万元Pre-A轮融资，投资方包括财通资本和商汤国香。钱不算多，但方向值得关注。这家公司不做机器人本体，而是专注做“大脑”，一个能融合图像、点云、超声等多种数据，对混凝土开裂、钢结构锈蚀等病害进行亚毫米级定量识别的工程多模态大模型。换句话说，它想让机器人具备“工程师”的思维，而不只是“巡检员”的腿脚。舒江鹏直言，行业痛点就是“强巡弱检”，机器人看到裂缝后还得靠人肉眼复核，这简直是对AI的侮辱。

技术亮点在于数据壁垒。旷行号称拥有全球基建领域最大的结构化多模态数据库，涵盖数百万个专业标注负样本，比同行高出两个数量级。这听起来很唬人，但背后是浙大土木系十五年的积累。在AI领域，数据就是石油，而垂直领域的专业数据更是稀有燃料。其他公司如果只靠公开数据集训练模型，在诊断精度上很难不被碾压。不过，这里有个潜台词：工业场景的负样本数据采集本就艰难，旷行靠学术背景挖到这座金矿，其他团队想追赶恐怕得从头挖井。

商业化路径上，旷行选择轻资产模式——采购或合作生产机器人本体，加装自研的“大脑盒子”。这挺聪明，避开硬件制造的重资产陷阱，专注软件和算法输出。但风险也很明显：本体厂商如果自己补足AI能力，旷行的护城河还剩多宽？目前看，它和江铜、国网等头部企业签了数千万元合同，矿下四足机器人已在试点。舒江鹏算的经济账很实际：一台机器人成本几十万，替代年成本超百万的人工，还避免停产风险。这逻辑成立，但前提是机器人足够可靠。他自己也承认，当前覆盖率只有50%，剩下一半还得靠人工。从50%到90%的提升，绝不是靠迭代算法就能轻松跨越——矿山井下网络波动、巷道微调，都可能让机器人“犯傻”。具身智能的战场从来不在实验室，而在泥泞的现实里。

关于人形机器人的争论，舒江鹏态度务实：短期主攻四足，因为稳定性更高，摔倒损失在高危场景中无法承受。这给当下一股脑追逐人形机器人的热潮泼了盆冷水。工业领域需要的不是外形像人的花瓶，而是能啃硬骨头的工具。人形机器人或许未来会有用武之地，但指望它们三五年内下矿井检修，无异于痴人说梦。旷行明确表示本体不是重点，只做“大脑”，这种聚焦值得肯定——毕竟术业有专攻，AI公司非要去和硬件厂商抢饭碗，往往死得难看。

投资方的观点也耐人寻味。商汤国香看好“空-地-矿一体”的系统方案，财通资本则强调“手-脑-眼”协同。但抛开这些术语，本质都是赌具身智能能在传统工业中撕开缺口。特种机器人替人趋势不可逆，尤其是年轻人宁愿送外卖也不愿下矿井的今天。不过，投资热潮中常有泡沫：多少公司拿着漂亮PPT融钱，最后连个稳定运行的案例都拿不出。旷行能跑到订单落地，至少证明技术不是空中楼阁。

总体而言，旷行科技的思路是对的：工业机器人不能只当移动传感器，得有诊断和干预能力。但前路依旧荆棘密布。数据积累需要时间，算法泛化面临场景刁难，商业化还得跨过成本与信任的门槛。舒江鹏说“做最懂工程的机器人大脑”，这话听着提气，但工程世界的复杂程度远超代码逻辑。或许，这场突围战才刚刚开始，而我们需要更多这样的公司，用扎实的数据和精准的痛点切入，把AI从幻象拉回泥土。