智东西（公众号：zhidxcom）
作者 | 周加琦
编辑 | 漠影
机器人的手正在变得越来越灵活。
抓取物品、拧螺丝、操作工具……越来越多过去需要人手完成的动作，如今灵巧手也能做到。
但真正进入工厂，情况就不一样了。
目前，工业自动化仍主要依靠夹爪完成特定工序，灵巧手的应用也多集中在单项操作。如何让同一只灵巧手适应不同工具，连续完成一整套工作流程，仍然是一项挑战。
智东西10月29日报道，今天，在智元创新日灵巧操作分论坛上，临界点现场连线卧龙希尔工厂，展示其最新工业POC成果：通过“机械臂+灵巧手”的组合，连续完成包含72个步骤的灵巧手性能测试流程。
这套流程涉及上下料、扫码、抓取螺丝枪、拧卸螺丝、按压按钮等多种操作。一期验证数据显示，任务成功率达到98%，单次作业仅需8.5分钟，已经接近人工约8分钟的作业节拍，并可连续运行8—12小时。
同时，临界点还推出了全新遥操作与数据采集一体手套OmniGlove Pro，能够同步采集人手动作与触觉信息，并已完成操作数据在机器人真机上的Replay验证。
此次工业POC与数采新品的亮相，也进一步展现了临界点围绕灵巧手本体、数据、模型与实际部署构建的全链路技术布局。
一、72步连续操作，作业效率接近人工
在卧龙希尔的灵巧手性能测试工位上，一只灵巧手正在给另一只灵巧手做测试。
仔细看，两只手的分工完全不同。安装在机械臂末端的手，负责完成整套操作；传送带上等待检测的，则是临界点即将出厂的灵巧手产品。
临界点模型算法总监吴铭东表示，临界点正在用自家的灵巧手，测试自家的灵巧手。
机械臂带动末端灵巧手抓取待测“手”，将其放置在检测设备上，随后扫码识别信息。紧接着，灵巧手抓起旁边的螺丝枪，对准螺丝孔，将待测产品固定在检测工位。放回工具后，机器人按下启动按钮，开始进行性能测试。
测试完成后，灵巧手又会按下停止按钮，再次拿起螺丝枪，卸下固定螺丝，将待测产品放回传送带。
这套流程涉及待测灵巧手的柔性装配、使用螺丝枪拧卸螺丝，以及测试设备的仪表盘操作三类不同操作，分别考验了其复杂物体抓取与定位、精细力控及工具操作、按键交互能力等能力。
整套流程包含72个操作步骤，全部由一台机械臂搭配一只灵巧手自主完成。
据现场介绍，临界点之前尝试过使用夹爪，但拧螺丝时的反作用力容易导致螺丝枪脱手。如果将工具固定在机械臂末端，又无法兼顾抓取产品、操作仪表盘等任务。因此，临界点选择使用五指灵巧手，无需更换末端，就能连续完成不同类型的操作。
吴铭东透露，工位一期POC的任务成功率达到98%，机器人最长能够连续不间断作业12小时。
并且机器人全自主完成整套72步测试流程仅需8.5分钟，而人工完成同样的工作约需8分钟。两者仅相差30秒，机器人单次作业耗时比人工仅高出约6.25%。
这意味着，一台机械臂搭配一只灵巧手，已经能够以接近人工的作业节拍，连续完成抓取、装夹、拧卸螺丝及设备操作等不同任务。相比直接部署整台人形机器人，采用“机械臂+灵巧手”的组合，也有助于控制设备投入成本。
据了解，临界点计划于年内推动该方案正式并线上岗，进一步检验其在真实生产中的长期稳定性和部署成本。
对于这一应用场景，智元副总裁、临界点CEO乔天杰的观点是，一双手在产线上装配着下一双手，这就是临界点的未来。
二、亚毫米级精准对位，受干扰还能自主恢复
这些看似普通且容易的操作，真正交给机器人执行，却有不少技术难点。
首先是螺丝枪的精准对位。在实际操作中，螺丝枪批头需要准确进入螺丝帽，对位精度要求达到亚毫米级。即使是经验丰富的工人，也需要不断调整工具的位置和角度，才能确保准确对位。
为了重复测试，工位使用的是可反复使用的防脱螺丝，螺丝在孔位中会产生毫米级的晃动，进一步增加了对位难度。
与此同时，螺丝枪本身较大、较重，连接线缆也会产生牵拉，拧紧时还有反冲力。即使灵巧手握持螺丝枪的姿态只偏差1毫米，也可能导致枪头偏移3—4毫米。而在此次工业POC中，临界点的灵巧手已经能够自主完成这一精细操作。
在现场，临界点还通过人为制造干扰，展示了机器人的自主纠偏和失败恢复能力。
工作人员在机器人拧螺丝时推动机械臂，使螺丝枪偏离原本的位置。可以看到，机器人并没有停止作业，而是自主调整位置和姿态，重新找到螺丝孔，继续完成拧紧操作。
另一项测试中，工作人员人为调整工装底座，使螺丝孔位发生偏移，导致螺丝只拧入一部分。机器人通过拧紧力和视觉信息识别出操作失败，随后重新对准孔位，再次拧紧螺丝，最终完成任务。
为了应对这些复杂情况，临界点采用了真机强化学习与智能体系统相结合的方案。
其中，强化学习让机器人能够根据视觉、触觉反馈实时调整动作，提高精准对位和抗干扰能力。临界点披露，经过真机强化学习训练后，机器人拧螺丝的速度比遥操作快4倍以上，对位动作也更加直接。
对于已经发生的操作失败，临界点则通过工业部署态智能体框架Omni-Maestro进行处理。该系统能够结合视觉、力矩及工位设备反馈判断任务是否完成，并在出现异常时重新执行，或回溯至出错步骤进行修复。
三、补齐高质量数据，加速“部署态”规模化
要让灵巧操作走向更多工业场景，推动“部署态”真正落地，高质量操作数据的获取是关键一环。
目前，Ego数据是具身智能规模化训练的一条重要路径。相比依赖机器人真机采集，通过记录人类日常操作获取数据，更容易扩大数据规模。
但传统Ego数据也有局限。这类方案主要依靠摄像头记录人手动作，再通过算法估计手部姿态，容易受到视觉遮挡等因素影响。同时，由于缺少触觉信息，也难以还原人手与物体接触时的状态。
针对此，临界点团队自研了数据采集手套OmniGlove Pro，并同步开启量产交付。
这款手套能够同时采集人手的位置、姿态及触觉信息，将其统一打包输出，为后续机器人操作提供数据支持。
具体来看，在动作采集方面，OmniGlove Pro每个指尖都有EMF磁场定位和IMU抗干扰融合技术，定位距离精度可达0.1mm！在触觉采集方面，这款手套配置了746个触觉感知点位，空间分辨率可达3mm，能够更准确地记录手部与物体的接触情况。
这些数据经过处理和跨本体动作重定向后，能够在机器人真机上Replay，完成与人类示范相同的任务。
临界点首席具身科学家黄碧丹将这类数据称为“真机Ready”的Ego数据。
这些数据，既有Ego数据的规模化优势，也有真机数据的高质量，更有人类操作的真实手感。三者结合，我们有望看到灵巧操作规划化部署进程的加速落地。
结语：灵巧手走向工厂，用实际表现说话
此次工业POC与OmniGlove Pro的发布，分别对应临界点在灵巧操作落地中的两个关键环节：前者通过真实工位验证灵巧手连续完成复杂任务的能力；后者则让人类操作数据能够更高效地用于机器人真机，进一步衔接数据采集与实际操作。
随着灵巧手逐步进入工业场景，能否在复杂工况下保持操作精度、以接近人工的节拍稳定作业，并将这些能力复制到更多场景，将成为其规模化应用的关键。
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。
Intellectual (WeChat Official Account:zhidxcom)
Zhou Ga Qi
Edit Queen
The robot's hands are getting more flexible.
Grab objects, screws, operating tools... More and more of the moves that used to be done by people can be done by smart hands.
But real access to the factory is different.
Currently, industrial automation continues to rely mainly on claws for specific processes, and smart-hand applications are concentrated in individual operations. The challenge remains how to adapt the same smart hand to different tools and to complete a set of workflow courses in succession.
On October 29th, Ji-King reported that today, at the Smarter Working Sub-forum on Intellectual Capital Innovation, the critical point was connected to the Dragonhill Plant to showcase its latest industrial POC results: through a combination of “mechanized arms + smart hands”, the smart hand performance testing process, comprising 72 steps, has been completed in succession.
The process involves a variety of operations, such as top-downs, sweeps, grabs of screwdrivers, screws, press buttons, etc. The first validation data showed that mission success rate had reached 98 per cent, that it took only 8.5 minutes for a single operation, that it was close to a manual work beat of about 8 minutes and that it could run for 8-12 hours.
At the same time, the critical point also introduced a new remote-operated and data-collective OmniGlove Pro, which is capable of synchronizing the collection of human action and tactile information, and has completed the validation of the operational data for Replay on the robot real machine.
The presentation of the industrial POC with the new product further demonstrates the full-link technology layout of the tipping point around the smarthands, data, models and actual deployments.
One, 72 steps of continuous operation. Operational efficiency close to manual.
On the smart hand test site at Dragon Hill, one smart hand is testing the other.
Carefully, the division of labour between the two hands is completely different. Hand installed at the end of the arm of the machine to complete the full set of operations; the conveyor belt, waiting for detection, is a smart hand product at a critical point about to leave the plant.
The Director-General of Critical Point Model Algorithms, Wu Songdong, stated that the critical point was being tested with his own smart hands.
The mechanical arm drives the end-of-pipe smart hand to the "hand" to be measured and places it on the detection device and then scans the identifying information. Immediately thereafter, smart hands grab the screwdriver next to it, point to the screwdriver, and hold the product to the test site. When the tool is put back, the robot presses the start button and starts the performance test.
Upon completion of the test, the smart hands will press the stop button, pick up the screwdriver again, unload the fixed screw and put the product to be tested back on the conveyor belt.
The process involves three different types of operation, namely, flexible assembly by a smart-tester, screw-throwing screws with a screwdriver, and dashboard operations for test equipment, which test the ability of complex objects to capture and locate, fine control and tooling, and key interaction.
The process package consisted of 72 operational steps, all of which were done by a mechanical arm with a smart hand.
According to a live introduction, the use of claws had been tried before the tip point, but the back-effects of screwdrivers could easily lead to screwdrivers. If the tools are fixed at the end of the mechanical arm, it is not possible to balance the tasks of capturing the product, operating the dashboard, etc. Therefore, the critical point chooses to use a five-fingerer, without the need to replace the end end, so that different types of operations can be performed in a continuous manner.
Wu Zhentung revealed that the task of POC phase success rate was 98 per cent and that robots could operate for up to 12 hours without interruption.
And it takes only 8.5 minutes for the robot to complete the full 72-step test process, while it takes about 8 minutes for the same work to be done manually. The difference between the two is only 30 seconds, and a single operation by a robot takes only about 6.25 per cent more time than manual.
This means that a mechanical arm combined with a smart hand has been able to perform different tasks, such as capturing, packing, wringing screws and equipment operation, on a continuous basis with close to a manual operational beat. The direct deployment of a whole-man robot using a combination of “mechanical arm plus smart hand” also helps to control the cost of equipment input.
It is understood that the Critical Point plan will promote formal and online induction of the programme during the year to further test its long-term stability and deployment costs in real production.
For this application, Vice-President Jiwon, CEO of the Critical Point, Joe Tianjie, is of the view that the future of the critical point is one hand on the production line and one hand on the production line.
ii. Sub-millimetre precision. Interrupted and autonomous recovery
These seemingly ordinary and easy operations, which are actually carried out by robots, have a lot of technical difficulties.
First is the precision of the screwdriver. In practice, the screwdriver batch needs to be accurately entered into the screwdriver, and the degree of precision of the position requires sub-millimetres. Even experienced workers need to constantly reposition the tool and its angle to ensure its accuracy.
In order to repeat the tests, the position uses reusable detachable screws, which generate millimetre-scale swings in the hole, which further complicates the position.
At the same time, the screwdriver itself is large and heavy, and the connection cable also creates a pull, and when it is tightened, there is a backlash. Even if the smart hand held the screwdriver had a single millimetre deviation, it could result in a shift of 3-4 mm in the head. In this industrial POC, the smart hands of the tipping point have been able to do this fine operation on their own.
On the ground, the tipping point has also demonstrated the ability of robots to recover autonomously and through failure by creating artificial interference.
The staff pushes the mechanical arm when the robot screws, diverting the screwdriver from its original position. As you can see, the robots did not stop working, but managed to adjust their positions and positions, rediscover the screwdrivers, and continue to do this.
In another test, the staff man-made retrofitting of the floor seat caused a shift in the screwhole, which resulted in the screw being only partially sprouted. The robot identifies the operational failure by wringing and visual information, then re-directs to the hole, and then tightens the screws and finally completes the mission.
To address these complexities, the cut-off points have adopted a combination of real-time enhanced learning and AI agent systems.
In this context, enhanced learning enables robots to adjust their actions in real time, based on visual and tactile feedback, so as to improve their accuracy and resistance to interference. Critical points reveal that robot screws are more than four times faster than remotely operated and are more direct in position after intensive learning.
For operational failures that have occurred, the critical points are processed through Omni-Maestro, the industrial deployment-type AI agent framework. The system is able to determine whether the task has been completed in conjunction with the feedback from visual, force-tangular and position-based devices, and is re-executed in the event of an anomaly or is retrospected to the wrong step.
Completing quality data to accelerate the “deployment” scale
Access to high-quality operational data is key to moving smart operations towards more industrial scenarios and to moving “deployance” to real ground.
Currently, Ego data is an important path to smart scale training. It is easier to scale data than to rely on robotics to collect data by documenting human daily operations.
But traditional Ego data also have limitations. Such programmes rely mainly on cameras to record manual movements, and can then estimate hand postures by algorithms, which are vulnerable to such factors as visual masks. At the same time, lack of tactile information makes it difficult to restore the state of human hands in contact with objects.
In response, the Critical Point team developed the data collection glove OmniGlove Pro and synchronized delivery.
The glove is capable of collecting both the position, attitude and tactile information of the person, and of delivering it in a single package, providing data support for subsequent robotic operations.
Specifically, in terms of action collection, OmniGlove Pro has EMF magnetic field positioning and IMU anti-disruption integration technology at every fingertip, with a location distance of 0.1 mm! In the area of tactile collection, the glove is equipped with 746 tactile points and space resolution can be as high as 3 mm, allowing for a more accurate recording of hand contact with the object.
The data, processed and re-directed across the body, can be used to replay on robotics and perform the same tasks as humans.
The lead critical point body scientist, Huang Beadan, referred to such data as “Ready Real Machine” Ego data.
These data range from the size of the Ego data to the high quality of the real data and a real sense of human hands. Together, we are expected to see the accelerated landing of a smart operational planning deployment process.
Conclusion: Smart hands to the factory, speaking with real expression
The industrial POC and OmniGlove Pro publish, respectively, correspond to two key points of the tipping point in a smart operation: The former validates the ability of smart hands to perform complex tasks on a continuous basis through real jobs, while the latter allows human-operated data to be used more efficiently in robotic machines and further links data collection and operations.
As smart hands move into the industrial landscape, the ability to maintain operational precision under complex conditions, close to artificial rhythm stabilization operations, and to replicate these capabilities to more scenes, will be key to their scale application.
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.