system
Before a robot can move, grasp, or plan anything, it first has to sense the world accurately, through vision, touch, and knowing its own position. Every other system depends on getting this right first.
视觉、激光雷达、触觉、融合与状态估计——所有下游的前置条件。
Bottlenecks瓶颈 P01 Multimodal sensing & calibration多模态传感与标定 · P02 Tactile & force sensing触觉与力感知 · P03 Fusion & state estimation融合与状态估计
What it isSenses is how a robot perceives the world: the layer everything else depends on. First, raw hardware captures the physical world: cameras and sensors that see and measure what's around the robot (P01, Multimodal sensing & calibration). Second, because vision and distance-sensing alone can't tell a robot how hard it's gripping something or whether it's about to drop it, touch-based sensing fills that gap (P02, Tactile & force sensing). Third, all of these raw, noisy signals get combined into one reliable picture of where the robot and the world actually are (P03, Fusion & state estimation). These three build on each other in order: sense, then add touch where sight can't reach, then combine it all into one answer. A weak link anywhere degrades everything built on top of it.
是什么“感官”（Senses）是机器人的感知技术栈：捕捉物理世界的传感与标定硬件（P01，多模态传感与标定）、视觉与测距都无法替代的接触式感知（P02，触觉与力感知），以及把这些原始、带噪声的信号流融合为机器人与世界当前状态的单一可执行判断的融合与状态估计（P03，融合与状态估计）。三者处于同一条因果链上——先是信号转换，再是一种独立的接触模态，最后是推断——链条中任何一环薄弱，都会拖累建立在其上的一切能力。
ThesisSenses ranks lowest of all nine systems on CE-N: the weakest capital-efficiency position on the entire board. Yet it also contains the two best risk-adjusted public stock returns on this site (the best return once you account for how much risk was taken to earn it): Hesai (+19.35%) and RoboSense (+10.80%), both LiDAR makers (sensors that measure distance using laser light). That contradiction is exactly what this thesis has to explain. The market is happy to pay for cheap sensor cash flow; what it won't pay for is the underlying thesis, because a cheap valuation on a business that just makes sensors (a 'transducer business') isn't the same thing as cheap earnings. P02, Tactile & force sensing (touch and pressure sensing) was supposed to be this layer's scarce, defensible edge: contact feedback, not visual resolution, is what actually blocks a robot from handling objects precisely. But a widening field of funded competitors is now pricing that scarcity away, on the same trajectory that already collapsed LiDAR prices.
论点“感官”是当前 CE-N 再排名中得分最低的系统，却也承载着本表风险调整后回报最好的两个上市标的——禾赛（Hesai，+19.35%）与速腾聚创（RoboSense，+10.80%）——这正是这个论点必须解释的张力所在。市场乐于卖给我们便宜的传感器现金流；它不愿意卖给我们的，是这个论点本身，因为传感器业务上便宜的估值倍数，并不等于便宜的盈利。P02，触觉与力感知 本应是这一层稀缺、具有防御性的优势——阻碍灵巧操作的是接触反馈，而非视觉分辨率——但越来越多获得融资的竞争者，正沿着当年压垮激光雷达价格的同一条轨迹，把这份稀缺性也定价殆尽。
Products mapped to this system映射到本系统的产品
Flexible tactile sensors and electronic skin used for contact and grip feedback, spanning dedicated tactile-sensor makers already tracked on the site plus an e-skin line built as a subsidiary of a broader instrumentation business, alongside additional names from the conference doc not yet tracked on the site.
7 companies 7 家公司 1 manufacturers 1 家制造商 +6 named in conference materials, unverified +6 项来自会议材料，未经核实
Six-axis force/torque sensors and force/position sensing used for contact-rich manipulation and collision detection, built by on-site entities whose own component text names force or force/position sensing, plus dedicated F/T sensor makers named only in the conference doc.
1 companies 1 家公司 3 manufacturers 3 家制造商 +4 named in conference materials, unverified +4 项来自会议材料，未经核实
LiDAR, depth, stereo and industrial/machine-vision hardware and perception software used for range-finding, 3D mapping and inspection, spanning on-site LiDAR and vision companies plus additional names from the conference doc not yet tracked on the site.
10 companies 10 家公司 2 manufacturers 2 家制造商 +7 named in conference materials, unverified +7 项来自会议材料，未经核实
Optical and inertial motion-capture systems and data gloves used to generate proprioceptive ground truth for teleoperation and training data; entirely drawn from the conference doc, since none of these companies were found tracked anywhere else on the site.
+8 named in conference materials, unverified +8 项来自会议材料，未经核实
Haptic feedback devices, EMG sensing and XR tracking hardware used to capture human input and teleoperation signals; entirely drawn from the conference doc, since none of these companies were found tracked anywhere else on the site.
+7 named in conference materials, unverified +7 项来自会议材料，未经核实
General-purpose sensing chips and data-acquisition hardware that sit underneath many robotics instrumentation stacks; entirely drawn from the conference doc, since neither company was found tracked anywhere else on the site.
+2 named in conference materials, unverified +2 项来自会议材料，未经核实
Companies mapped to this system映射到本系统的公司
Manufacture mapped to this system映射到本系统的制造
Hesai’s FY2025 net income of RMB436m implied a 14.4% net margin and anchored the board’s best risk-adjusted public return. Roughly 40% of that income came from a single quarter’s gain on disposing of an equity stake — a non-operating item — and the very next quarter, Q1 2026, printed a net margin of just 2.6% on revenue that was still growing 30%.
RoboSense’s history is similar in shape: its first-ever quarterly profit landed in Q4 2025, after a full-year loss of RMB145m. Both companies are real and growing — Hesai’s robotics-and-ADAS lidar shipments and RoboSense’s robotics-lidar volume (up 1,458.8% year over year in Q1 2026, exceeding its own ADAS shipments for the first time) are genuine businesses — but the recurring earnings power behind the sensing sleeve’s headline returns is considerably thinner than the FY2025 net-income lines suggest.
The P02, Tactile & force sensing bottleneck the taxonomy originally called scarce has stopped looking scarce on the supply side in the space of about five months.
At least ten funded competitors are now shipping or claiming volume: PaXini raised roughly RMB2bn across two rounds in 2026 and claims sub-RMB-200 unit pricing; Tashan Technology raised three rounds in six months and claims tens of thousands of units shipped monthly; Daimon Robotics, Hanwei Technology, and at least six other named Chinese entrants are all shipping or scaling, alongside non-China entrants Tacta Systems, GelSight, XELA, and Contactile.
That is a commodity-component race in its second year, not a scarcity play — it is the same shape lidar had in 2021, before Hesai’s and RoboSense’s own ASPs collapsed from roughly RMB11,000 to roughly RMB1,660 per unit.
The lidar ASP-collapse-into-robotics thesis is the most corroborated claim in a batch of 20 sell-side reports: four separate houses (UBS, Deutsche Bank, China Renaissance, Guotai Haitong) each initiate or reiterate Buy ratings on Hesai, RoboSense, or new China names (Seyond, LDRobot) explicitly on the robotics-diversification story, even as ADAS pricing keeps compressing. UBS’s own move is the clearest tell: it cut its Hesai and RoboSense price targets on pricing-pressure earnings downgrades while holding both at Buy — “shrink the position, keep the thesis,” the exact move the primary evidence above already argues for.
On tactile, an independent Korean sell-side note — sourced separately from the Chinese press coverage already cited — corroborates the same commoditisation trajectory from the opposite side of the world: tactile-sensor unit cost falling from roughly $10,000 to roughly $350, a 95% decline.
J.P. Morgan separately flags six-axis force sensors, alongside planetary roller screws, as under 50% localised in China — a distinct, force-sensing-specific supply-chain gap.
On where the value actually sits, Morgan Stanley’s own numbers make the case better than the primary evidence alone: an L4 Waymo robotaxi carries roughly $30,000 of sensors and compute against a Tesla FSD stack at roughly $3,000, a 10x gap for the identical fusion and state-estimation problem, and the same house’s Humanoid Tech report argues the stack’s value sits in the AI brain, vision, and analog sensing chips — not the integrators. None of this changes the existing ranking; it corroborates “value migrates to fusion, calibration, and software, not raw transducers” from a source entirely independent of the Hesai, RoboSense, and Cognex reporting already in the primary evidence.
The moat in senses was never the transducer, and the evidence now says so three times over — from the primary earnings data, from the tactile cost-curve collapse, and independently from the sell-side’s own sensor-cost breakdowns. Cognex is the clean counterexample — eight consecutive quarters of revenue growth and margin expansion, a 32.2% adjusted EBITDA margin, on an installed base and qualification process that competitors cannot easily replicate — while raw sensor pricing (lidar, and now tactile) keeps collapsing toward commodity.
That argues for weighting incumbents with installed-base and qualification moats (Cognex, Hikvision/Hikrobot, and by extension whichever machine-vision names have real switching costs) over component innovators, and for treating Hesai’s and RoboSense’s headline net-income lines as lower-quality than their revenue growth implies until several more quarters confirm the recurring margin.
禾赛 2025 财年净利润 4.36 亿元人民币，隐含净利率 14.4%，是本表风险调整后回报最好的上市标的所依托的数字。其中约 40% 的利润来自单一季度的一笔股权处置收益——这是一项非经营性项目——而紧接着的下一个季度，即 2026 年第一季度，在营收仍同比增长 30% 的情况下，净利率仅录得 2.6%。
速腾聚创的走势形态类似：其历史上首个季度盈利出现在 2025 年第四季度，而在此之前是全年 1.45 亿元人民币的亏损。两家公司都是真实且在增长的企业——禾赛的机器人与 ADAS 激光雷达出货量、速腾聚创的机器人激光雷达出货量（2026 年第一季度同比增长 1458.8%，首次超过自身的 ADAS 出货量）都是真实的业务——但支撑这一传感器板块头条回报的经常性盈利能力，远比 2025 财年净利润数字所暗示的要薄弱得多。
本分类法最初称之为稀缺的 P02，触觉与力感知 瓶颈，在大约五个月的时间里，在供应端已不再显得稀缺。
目前至少有十家获得融资的竞争者在出货或宣称已具备量产能力：帕西尼（PaXini）在 2026 年两轮融资中共募集约 20 亿元人民币，宣称单价已降至 200 元人民币以下；他山科技（Tashan Technology）六个月内完成三轮融资，宣称月出货量已达数万台；戴盟机器人（Daimon Robotics）、汉威科技（Hanwei Technology）以及至少另外六家具名的中国新进入者都在出货或扩产，此外还有非中国厂商 Tacta Systems、GelSight、XELA 与 Contactile。
这是一场进入第二年的商品化组件竞赛，而非稀缺性叙事——其形态与 2021 年的激光雷达如出一辙，彼时禾赛与速腾聚创自身的平均售价随后从约 1.1 万元人民币崩跌至约 1660 元人民币。
激光雷达平均售价崩塌、转向机器人业务的论点，是 20 份卖方报告样本中印证最充分的一条：四家不同机构（瑞银、德意志银行、华兴证券、国泰海通）分别对禾赛、速腾聚创或新覆盖的中国标的（Seyond、LDRobot）发起或重申买入评级，明确以机器人业务多元化作为理由，即便 ADAS 定价仍在持续压缩。瑞银自己的操作最能说明问题：它在下调禾赛与速腾聚创目标价（因定价压力导致盈利预测下修）的同时，依然维持两者买入评级——这正是上文一手证据所主张的“缩减仓位、保留论点”的操作。
在触觉方面，一份独立的韩国卖方研究笔记——信源与前述中国媒体报道各自独立——从地球的另一端印证了同一条商品化轨迹：触觉传感器单价从约 1 万美元降至约 350 美元，降幅达 95%。
摩根大通则另外指出，六轴力传感器与行星滚柱丝杠一样，在中国的国产化率不足 50%——这是一个独立的、专属于力传感的供应链缺口。
在价值究竟落在何处这一问题上，摩根士丹利自己的数字比一手证据本身更有说服力：一辆 L4 级 Waymo 无人出租车携带约 3 万美元的传感器与算力，而特斯拉 FSD 方案仅约 3000 美元，对同一个融合与状态估计问题而言存在 10 倍差距；同一家机构的人形机器人科技报告认为，技术栈的价值落在 AI 大脑、视觉与模拟传感芯片上，而非系统集成商。这些都没有改变现有排名；它们从一个与禾赛、速腾聚创、康耐视这些一手证据完全独立的信源，印证了“价值向融合、标定与软件迁移，而非原始传感器”这一判断。
感官层的护城河从来都不在传感器本身，如今这一点已被证据三次印证——来自一手盈利数据，来自触觉成本曲线的崩塌，以及独立地来自卖方自身对传感器成本的拆解。康耐视（Cognex）是干净的反例——连续八个季度营收与利润率双双扩张，经调整 EBITDA 利润率达 32.2%，依托竞争对手难以轻易复制的存量装机与认证流程——与此同时，原始传感器（激光雷达、如今再加上触觉）的定价却在持续向商品化崩塌。
这意味着应更多配置拥有存量装机与认证护城河的龙头企业（康耐视、海康威视/海康机器人，以及其他任何拥有真实转换成本的机器视觉公司），而非组件层面的创新者；也意味着在有更多季度确认经常性利润率之前，应将禾赛与速腾聚创的头条净利润数字，视为质量低于其营收增速所暗示的水平。