FlexiTac压阻式传感器可兼顾柔性、低成本与一致性。
作者丨向欣
编辑丨岑峰
机器人的视觉能力正快速增强，但触觉仍然是具身智能系统中的薄弱环节。尤其是在精细操作场景中，当物体被遮挡、透明、易碎，或需要精确判断接触力与手内姿态时，仅靠视觉往往难以确认机器人是否真正抓稳了物体、是否发生了滑移，也难以实现可靠的力控与顺从交互。与图像领域已经形成相对统一的传感器和数据体系相比，触觉研究尚未形成类似的「收敛」。不同实验室往往使用不同类型的触觉传感器，但它们都面临寿命、柔性、多维信息感知、分辨率、成本与一致性难以同时兼顾的困境。这使得触觉数据采集、算法研究以及不同系统之间的复现都变得更加困难。如何让触觉像视觉一样，成为一种可以被更多机器人使用、被大量采集、被算法反复利用的基础感知能力？针对这个问题，哥伦比亚大学李昀烛团队提出了开源柔性压阻式触觉传感器平台 FlexiTac。在本届 IROS 2026 的触觉专题研讨会上，李昀烛系统介绍了这一平台开源一年半后的进展与生态。FlexiTac 是一套低成本、开源、可扩展的压阻式柔性触觉传感器，团队同时开源了完整 BOM、详细组装说明以及用于读取触觉数据的软件。FlexiTac 采用 FPC 与力敏薄膜层叠结构，组装仅需 3 分钟，可弯曲、裁剪并包裹于不同几何表面，适配手套、指尖或机器人皮肤等多种形态。其核心优势在于成本低、制作简单、高度柔性与可扩展、以及较好的长期稳定性和批次一致性，感知机制也足够简单，易于用弹簧-阻尼系统实现高保真仿真。其技术栈可集成进主流的机器人学习框架中，门槛低到拥有即使是本科生，只要拥有一台可打印 SO-100 机械臂的 3D 打印机，也能轻松把触觉传感器装上机器人。李昀烛提到，开源约一年半以来，FlexiTac 已在学术界与工业界获得广泛采用，衍生出多模态策略、可变形世界模型、四足移动-操纵、VLA 触觉反馈等多方向工作，他们与普渡大学合作的多指灵巧操纵工作 VTAP Gripper 还入围了本次 IROS 最佳论文奖。Analog Devices（ADI）正在推进其商业化版本，该版本触觉分辨率达到人类手指的五倍，基于半导体制造工艺量产。以下为李昀烛在 IROS 2026 演讲的精编稿，AI科技评论基于原英文内容进行了不改原意的翻译编辑：为了让传感器易于生产、成本低廉，我们最新的版本制造成本其实已经非常低，具体多低，取决于你同一批次生产多少片。准备好所有物料后，你只需要花 3 分钟就可以把传感器装配完成，传感器随即就能投入工作。我们这套方案是完全开源的：完整的物料清单，还有非常详尽的分步装配说明，同时也配套了可以读取、传输传感数据的对应软件。除了传感器和配套软件本身，我们也还把这套软硬件技术栈集成到非常流行的机器人学习框架 LeRobot 里。这样一来，任何人，哪怕只是本科生，只要手头有能够打印出低成本 SO-100 机械臂的 3D 打印机、都可以轻松地把我们的触觉传感器集成进去，然后开展相关研究，探索触觉感知如何帮助机器人理解手内物体状态，完成复杂操作任务。在我们开源这套传感器大约一年半之后，我们欣喜地看到，这项技术已经在学术界与工业界得到广泛应用。更让我们振奋的是，大家正在沿着不同方向继续推进、拓展整个生态：如何制造更好、能力更强的硬件；如何基于我们的触觉传感器构建更好的感知、学习和推理系统；以及如何设计真正包含触觉的大规模数据采集流程。我们也在与亚德诺半导体（Analog Devices，ADI）展开合作，由他们开发这款传感器的商业化版本。 他们已经在多场会议上演示了商业化样机，其中就包括 GTC 大会，演示内容是一些很有代表性的机器人操作任务。大家可以访问我们的项目网站，查阅硬件、软件方向的各类学术与产业应用成果。但话说回来，我们这款传感器只是众多触觉传感方案的其中一种方案。触觉领域常常被诟病的一个问题是，整个领域至今未形成统一标准。和发展成熟的计算机视觉不一样，不同实验室往往采用完全自研、互不兼容的触觉传感器。第一是可扩展性。它既可以做成大面积，也可以做成小尺寸。第二是柔性。它就像一张纸，可以弯折、裁剪，还可以包覆在各种异形曲面结构上。 第三是高度可定制化。我们内部已经完成原型验证，可以把传感器定制成触觉手套、指尖传感器、机器人皮肤等多种形态。第一个维度是时间维度：信号会不会随时间发生漂移，即长期的稳定性。这里举一个实例：我的学生 Binghao 主导了很多相关工作，他 2024 年夏天在实习期间，为录制视频制作了四块传感器。 时隔一年半之后，直接使用这套传感器，用一年前采集的数据训练出来的模型依然可以正常工作。第二个维度是不同批次制备之间的一致性：不同批次加工出来的传感器，输出信号能否保持一致。这是很多触觉传感器的痛点，尤其是结构中包含橡胶材料的传感器；而我们这款传感器在这一点上表现出不错的一致性。我们也正和亚德诺半导体合作，探索如何规模化生产这类传感器。我们可以很方便地把这款触觉传感器接入仿真环境，借助仿真做规模化算法迭代。下面我要介绍的后续工作正是基于这一点。因为传感机制简单，仿真实现只需要构建一个简单的弹簧‑阻尼系统，只需要标定 3 个超参数。左边是真实物理交互场景，后方屏幕可以看到触觉信号的实时可视化；右边是仿真环境，运行完全相同的策略，执行相同任务。 不仅状态演化过程高度吻合，仿真端屏幕输出的触觉信号也和真实信号匹配度很高。当然，你可能会说，这只是定性结果。那么定量结果表现如何？我们的做法是，先在真实环境中采集一批随机交互数据，然后把真实环境里的交互动作在仿真环境中复现回放。然后，我们分别收集真实环境和仿真环境得到的全部触觉数据，并绘制成直方图。标定工作的核心目标就是对齐这两组直方图，而二者实际对齐效果非常理想。这里我想特别强调一点：仿真与真实对齐得越精确，你在仿真中需要做的随机化就越少，sim-to-real 的迁移也会越精准。所以，在仿真环境中，我们可以进行大规模随机化，然后把训练出来的策略迁移回真实环境。同样的框架也可以应用到不同本体的机器人上。我们在实验室的双臂装配任务中进行了验证，也在英伟达的人形机器人上使用了同样的 sim-to-real 迁移框架，让机器人完成类似的任务。接下来，我想讲的是如何实现真实世界数据的规模化采集。很多人应该都熟悉 UMI 这套手持遥操作设备，它被用于野外场景的数据采集。用过 UMI 夹持器的人应该会注意到：设备内部的机械弹簧会给人类操作者提供触觉反馈，但原版 UMI 的设计只会记录精准的位姿跟踪信息，并不会存储这套触觉反馈。接入我们的触觉传感器恰好补上了这一短板。它可以间接把人类操作者感受到的触觉反馈，记录为传感器输出的触觉信号。借助这套改造后的手持设备，我们可以到各类不同场景完成大规模数据采集。采集到的数据可以同时包含视觉和触觉信息，而且两者是时间同步的。去过哥伦比亚大学校园的同学或许能认出其中部分场景。我们拿到视觉‑触觉配对的同步数据之后，可以用来做预训练，也有人称之为中间训练，取决于工作流的定义。之后，我们还可以继续采集遥操作数据，用于机器人策略的后续微调训练。这个案例是一项实验室自动化任务：机器人需要对试管做姿态重定向，过程中会有人为施加外部扰动。机器人必须持续调整试管姿态，才能够完成插入操作。 右侧是策略模型的输入，包含腕部相机图像以及触觉传感信号。经常有人提出一个问题：既然已经有腕部相机，我们还需要触觉吗？腕部相机的信息难道还不够用吗？ 这个任务就是很好的反例。如果试管是透明物体，仅依靠腕部相机，很难精确判断试管的姿态与位置；但下方的触觉信息就可以对机器人手内物体的状态做出精准估计。除此之外，我们还完成了很多其他任务：比如铅笔在手内姿态调整、把铅笔插入卷笔刀；还有需要精准力控的任务：控制夹持器张开吸取液体，但又不能张得过大导致玻璃器皿脱落；还有经典的恒力控制任务，比如白板擦拭。我们从一套已经发展数十年的简易传感机理出发，尽最大努力把这套传感器推广给更多研究者，我们把这件事情称为「让触觉感知大众化」。我们在努力地推进这件事情，看到这套传感方案被广泛采用，也由衷感到高兴。而在 AI 的世界里，我们也一直在思考如何通过仿真、真实物理环境两条路径，把触觉感知的规模进一步做大。而这一切仅仅只是开始。其中一项工作是与普渡大学的合作者共同完成的。我们不再只使用单一类型触觉传感器，而是把不同类型的触觉传感器结合起来，包括我们实验室的压阻式传感器，以及基于视觉的触觉传感器，并把它们集成到多指夹爪中，实现高灵巧度操作任务。这篇论文会在本次 IROS 上进行报告，同时也是最佳论文奖的入围作品之一。如果各位对该方向感兴趣，欢迎和我们交流。我们实验室另一项重要工作围绕世界模型展开。如大家所料，我们也在探索如何将触觉感知融入世界模型。大家现在看到的是基于动作条件的交互式 3D 视频预测，全程没有用到物理仿真，全部由真实数据训练得到。 这套模型可以输出帧率 15fps、同时包含视觉点云与触觉点云的预测结果。仔细看指尖部分可以发现：在三维空间中，触觉点云可以和物体底层几何结构实现很好的对齐。 模型还可以识别滑移：当两件物体对位不够好时，可以检测到滑移现象；当试管和支座对位精准，就可以完成插头插入试管的动作。触觉预测与视觉预测可以很好地协同对齐，这对于手内物体状态测量、估计这类高精度任务至关重要。
▎听众 A：我有一个关于触觉传感器切向力的问题。首先，你们的传感器能否测量切向力？ 其次，如果本身不做切向力测量，背后是什么原因？切向力对策略的成功是否重要？ 另外，在 UMI 数据集里，触觉传感器是如何生成动作标签的？ 如果去掉UMI内部的机械弹簧结构柔顺机构，触觉到动作标签的映射还能否正常工作？
第一个问题关于剪切力：压阻材料本身并不能直接测得剪切力。 但我们发现，如果传感器的空间分辨率、时间分辨率足够高，就可以对剪切力做近似估计。 我们亚德诺半导体的合作团队正在提升传感器分辨率，我记得目标分辨率至少是现有版本的 5‑7 倍。 他们发现在传感器表层覆盖一层极薄的软材料，材料发生形变之后，就可以实现对剪切力的近似估算。 读取触觉数据的时候，即便处于静态剪切状态，你也能看到力分布发生偏移；动态剪切当然也可以捕捉。所以我们认为，在传感器设计中引入剪切力测量拥有巨大研究潜力。他们已经有两篇相关论文被 CORAL 会议接收，会在 CORAL 做报告。其中一篇专门研究如何用压阻材料检测滑移。你的第二个问题，关于把传感器集成到 UMI 夹持器： 对于触觉可以发挥巨大价值的任务，尤其是需要柔顺特性、力级推理与力控操作的场景，触觉感知会起到至关重要的作用。这也是为什么我们的传感器和 UMI 夹持器兼容性很好：夹持器的手指本身具备柔顺特性，同时内置机械弹簧，给夹持机构带来柔顺能力。 我坚信，对于现实世界绝大多数有价值的机器人任务，这种柔顺特性都是必不可少的。谢谢。（雷峰网(公众号：雷峰网)）
为了方便大家抱团交流、互通会议消息，我们专门建了 IROS 2026 参会交流群！进群你会解锁这些「福利」?
- 会议情报站：投稿 DDL、格式规范、评审进度……关键节点第一时间送达，再也不怕错过 deadline，投稿准备稳稳的。
- 同行茶话会：天南海北的同行都在群里，聊心得、碰思路、攒人脉，科研路上我们同行。
- 前沿补给站：最新研究成果、热点方向实时同步，结合会议主题帮你捋清投稿脉络，给论文灵感加满油。
- 现场后援团：（会议期间专属开启）：到了会场也不用慌—— 路线导航：场馆分布、报告厅怎么走，群里随时问； 议题共读：热门 session、Keynote 探讨，错过也能追； Poster 汇编：现场海报精华整理，一键收藏不遗漏； 约局广场：约饭局、采访局、交流局……想唠嗑、想合作、想面基，群里发起就成。
? 进群传送门： 扫码进群或添加微信Luyoyo_2026，备注：ECCV + 单位/学校+姓名+方向。
搞科研/搞技术，信息差很重要。
来，一起快人一步！
可独家畅览：
专家演讲PPT
大会报告全文
热门论文解读
学术新星访谈
扫描上方二维码
或点击「阅读原文」关注专区。
雷峰网原创文章，未经授权禁止转载。详情见转载须知。
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
FlexiTac pressure-resisting sensors can combine flexibility, low cost and consistency.
The author is happy.
Edit Peak
The visual capacity of robots is increasing rapidly, but the touch is still a weak link in the smart system. In particular, in a fine operating scene, when objects are shielded, transparent, fragile, or require a precise judgement of their exposure and hand-held posture, it is often difficult to ascertain from vision alone whether robots really hold the object, have slipped, or have a reliable control and conformity interaction. As compared to the relatively uniform sensor and data systems that have developed in the area of imagery, tactile research has not yet developed a similar "diffusion". Different laboratories often use different types of tactile sensors, but they all face the dilemma of longevity, softness, multidimensional information perception, resolution, cost and consistency. This makes it more difficult to collect tactile data, to study algorithms and to reproduce between systems. How can touch be, like vision, a basic sensory ability that can be used by more robots, captured in large quantities and used repeatedly by algorithms? In response to this problem, the Li Candles Team at Columbia University proposed the Open Source Soft Pressure Interception Sensor Platform FlexiTac. During this IROS 2026 tactile workshop, Li candle system presented progress and ecology after a year and a half after the platform was opened. FlexiTac is a low-cost, open-source, scalable, flexible touch sensor with a full BOM, detailed assembly instructions and software for reading tactile data. FlexiTac uses FPC and a thin film layering structure, which takes only three minutes to assemble, to bend, tailor and wrap in different geometric surfaces, suitable for various forms such as gloves, fingertips or robot skin. Its core advantages are that it is low-cost, simple production, highly flexible and scalable, as well as better long-term stability and batch consistency, and that the sensory mechanisms are simple enough to achieve high-prevalence simulations with spring-train systems. Its technology warehouse integrates into the mainstream robotic learning framework, and the threshold is so low that even undergraduate students can easily load a tactile sensor with a 3D printer that can print SO-100 mechanical arm. Li Zhong mentioned that, for about a year and a half, FlexiTac had been widely used in academia and industry, producing multi-modal strategies, deformable world models, four-legged mobile-manipulation, VLA touch feedback, and that VTAP Gripper, who was working with the Pudu University, had also joined the award for the best IROS thesis. Analog Devices (ADI) is advancing its commercialized version, a tactile version of resolution five times the human finger, based on the semiconductor manufacturing process. The following is an excellent version of the Li Xiao Chong's speech at IROS 2026, and the AI Technological Review has been translated and edited on the basis of original English content: In order to make sensors easy to produce and cheap, the cost of making our latest version is already very low, depending on how many pieces you produce in the same batch. Once you've got all the material, it takes you three minutes to get the sensors together and then the sensors go into work. Our package is fully open-source: a complete inventory of materials and a very detailed step-by-step presentation, accompanied by matching software for reading and transmitting sensor data. In addition to the sensors and the software package itself, we have integrated the software and hardware technology into a very popular robotic learning framework, LeRobot. In this way, anyone, even undergraduates, can easily integrate our touch sensors with 3D printers that can print low-cost SO-100 mechanical arms. In Go and then conduct research to explore the sensory perception of how to help robots understand the state of their hands and perform complex operational tasks. About a year and a half after our open source set of sensors, we are pleased to see that this technology has been widely applied in academia and industry. We are encouraged by the fact that you are continuing in different directions to expand the whole ecology: how to create better and more capable hardware; how to build better sensory, learning and reasoning systems based on our tactile sensors; and how to design large-scale data-collection processes that really contain touch. We're also working with Analog Devices, ADI, to develop a commercial version of the sensor. They have demonstrated the commercialization of the prototypes in several sessions, including the GTC Congress, which is a very representative robotic operation. You can visit our project website and see the results of various academic and industrial applications in hardware, software orientation. But then again, our sensor is just one of many sensory programmes. One of the problems in the tactile field is that so far there has been no uniform standard in the whole area. Unlike the development of mature computer vision, different laboratories often use fully self-researched, incompatible tactile sensors. The first is scalability. It can be either large or small. The second is softness. It's like a piece of paper. It's bending, tailoring, and it's covered in a variety of alien arcs. The third is highly customized. Prototype validation has been completed internally, so that sensors can be customised into tactile gloves, finger-point sensors, robotic skin, etc. The first dimension is the time dimension: will the signal drift over time, i.e. long-term stability? Here's an example: my student Binghao, who led a lot of work, developed four sensors for video recording during his internship in the summer of 2024. After a year and a half, this sensor is used directly, and models trained with data collected a year ago are still working. The second dimension is consistency between different batch preparations: the sensors that are processed in different batches and the consistency of the output signals. This is the pain point of many tactile sensors, especially those that contain rubber materials in their structures; and we have a good consistency on this point. We are also working with the Adino semiconductor to explore how to scale up the production of such sensors. We can easily access this tactile sensor into the simulation environment, using simulation to scale algorithms. This is the basis for the follow-up that I would like to present. Because sensory mechanisms are simple, imitation only requires the construction of a simple spring-based resistance system and the marking of three super-parameters. The left is a real physical interactive scene and the rear screen can see real-time visualization of the tactile signals; the right is a simulation of the environment, running exactly the same strategy and performing the same task. Not only does the state evolution process correspond at a high level, but the tactile signal of the simulation-end screen output also matches the real signal very well. Of course, you might say it's just a qualitative result. How did the quantitative results work? Our approach is to collect a random set of interactive data in the real environment and then replay the interactions in the real environment in the simulation. Then we collect all the tactile data from the real environment and the simulation environment, and we draw a histogram. The core objective of the mapping exercise is to align the two sets of histograms, which are ideal for their actual alignment. Here I would like to emphasize in particular that the more accurate the simulation is, the less random you have to do in the simulation, sim-to-rea. I'll move more precisely. So, in a simulation environment, we can make large-scale randomizations, and then we can move our training strategies back to the real environment. The same framework can also be applied to robotics of different kinds. We've tested it in the lab's arms assembly mission, and we've used the same sim-to-real migration framework on the human robots in Weaverda to do similar tasks. Now, what I'd like to talk about is how to get real-world data to scale. A lot of people should be familiar with UMI, which is a remote operating device that is used to collect data on field scenes. Those who have used the UMI handler should have noticed that the mechanical spring inside the device will give human operators a touchback, but the original UMI design will only record accurate position tracking information and will not store this touchback. Access to our tactile sensors just fills this panel. It can indirectly record the tactile feedback felt by human operators as a sensor's output. With this modified handheld, we can go to a variety of scenarios to complete large-scale data collection. The data collected can contain both visual and tactile information, and both are time synchronized. Some of them may be recognized by students who have been to Columbia campus. When we get the visual-touch pair of synchronized data, it can be used for pre-training, and some people call it intermediate training, depending on the definition of workflow. And then we can continue to collect remote operational data for subsequent fine-tuning of robotics. This case is a laboratory automation mission: robots need to re-direct their attitude towards the test tubes, during which humans can impose external disturbances. The robot must continuously adjust the test tube attitude to complete the insertion operation. The right side is the input of the strategy model, which contains a wrist camera image and a tactile signal. A question is often asked: Do we need to be touched now that we have a wrist camera? Isn't that enough for the wrist camera? The mission is a good example. If the test tube is a transparent object, relying on a wrist camera alone, it is difficult to determine precisely the attitude and position of the test tube; however, the sensory information below provides a precise estimate of the condition of the robotic object. In addition to that, many other tasks have been accomplished: pencils are adjusted in hand, pencils are inserted into pencils; there are tasks that require precision control: controls are open to draw liquids, but not too large to cause glass to fall; and there are classic constant control tasks, such as whiteboard wipes. Proceeding from a set of simple sensory mechanisms that have been in place for decades, we do our best to extend this sensor to more researchers, and we call it "the popularization of tactiles". We are trying to move this forward, and we are pleased to see that this sensory package has been widely adopted. And in the world of AI, we've been thinking about how to take the sensory scale further through the two pathways of simulation, real physical environment. And it's just the beginning. One of these efforts was carried out jointly with partners at the Pudu University. Instead of using only single types of tactile sensors, we combine different types of tactile sensors, including pressure-resistant sensors in our laboratories, as well as visual-based tactile sensors, and integrate them into multi-finger claws to achieve high-intensity operations. This paper will be presented at this IROS and is also one of the winnings of the Best Thesis Award. If you are interested in that direction, you are welcome to communicate with us. Another important work of our laboratory revolves around world models. As you can imagine, we are also exploring ways to integrate sensory perception into world models. Guys. An interactive 3D video projection based on the conditions of action is now available, and the entire process is not used for physical simulation, all of which is obtained through real data training. This model can output 15fps with both visual and tactile cloud predictions. A closer look at the tip of the finger shows that, in three-dimensional space, the tactile cloud can be well aligned with the substrate geometry of the object. Models can also recognize slides: slides can be detected when the two objects are not aligned well enough; and action to insert a test tubes can be completed when the test tube and the feed are accurate. tactile predictions and visual predictions can be well aligned, which is essential for measuring the state of the objects on hand and estimating such high-precision missions.
A: I have a question about tactile sensors. First of all, can your sensors measure the trajectories? Secondly, what is the reason behind this if it does not take a cut-throat measure itself? Is it important for the strategy to be successful? Also, in the UMI data set, how do tactile sensors generate action labels? If the mechanical springs inside UMI are removed, can the mapping of action labels work properly?
The first question concerns shearing: the pressure material itself cannot be directly measured. But we found that if the sensor's space, resolution, time resolution, was high enough, a similar estimate of shearing could be made. We're working with the Yadno semiconductor to upgrade the sensor resolution, and I remember target resolution is at least 5-7 times the current version. They found a very thin layer of soft material on the surface of the sensor, and after the material had been shaped, similar estimates of shearing power could be achieved. When you read tactile data, you can see a shift in power distribution even when you're still cutting; dynamic cutting can certainly be captured. So we think there's great research potential for introducing shearing measurements into sensor design. They already have two related papers received by the CORAL meeting and will report to CORAL. One of them is dedicated to researching how to detect slides with pressure-resistant materials. Your second question, on integrating sensors into UMI binders: Sensory awareness plays a crucial role in the task where touch can be of great value, especially in situations that require softness, force-level reasoning and control. And that's why our sensors and UMI handlers are so compatible: the handlers' fingers are soft in themselves, and they are built into mechanical springs that give them a softer capacity. I firmly believe that this softness is essential for the vast majority of the valuable robotic tasks in the real world. Thank you. (Ray Peak Network (WeChat Official Account:Ray Peak Network))
IROS 2026 has been set up to facilitate exchange and exchange of information. You're gonna unlock these "freebies"?
- Conference intelligence posts: contribution DDL, format specifications, review of progress... Key nodes arrive at the first moment, no longer afraid of missing trackline, and the contribution is prepared to be steady.
- Tea sessions: We're all in a group of people who are in the North and the South China Sea, and we're all on the same path.
- Front depots: recent research, real-time synchronization of hotspots, to help you clear the text of the paper in conjunction with the theme of the conference, and fill the paper with oil.
- On-site support missions: (opened only during the session): no panic at the meeting venue - route navigation: the distribution of the premises, how the reporting room goes, and how the group asks;
♪ Into the stratum portal: Scanning in groups or adding WeChatLuyoo 2026, note: ECCV + unit/school + name + direction.
Research/technology, poor information is important.
Come on, hurry up!
It's an exclusive view:
Expert presentations PPT
* Reissued for technical reasons.
The reading of popular papers
Academic New Star Interview
Scan 2D above
This post is part of our special coverage Syria Protests 2011.
Original Ray Peak article, unauthorized reproduction. For further information, see the information below.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.