Ommo Technologies is building the hand data robots learn from

The next bottleneck in robot manipulation is showing up at the fingertip, not in another stage demo. According to a Chinese tech media report, Ommo Technologies chose permanent-magnet magnetic-field positioning after evaluating optical positioning and millimeter-wave radar, because mainstream optical positioning can fail when objects block the line of sight.

That detail points to a larger shift.

A hand that can grasp is useful. A hand that can measure what happens during the grasp is the start of a data loop. Ommo's route uses mechanical rotation of a permanent magnet to generate a characteristic magnetic field, then calculates complete 6DoF position and posture data from real-time signals collected by miniature magnetic sensors. Its smallest sensor is about 0.8 millimeters, and the system can currently reach sub-millimeter accuracy within a specified working range.

The pattern is clear: better embodied AI depends on treating hands, sensors, teleoperation tools, and data pipelines as one machine for producing repeatable manipulation evidence.

The hand is becoming a data instrument

Robot-hand debates commonly start with the visible mechanics: five fingers, tendon drives, modular fingertips, more joints. The more useful question is what the hand can record while it is touching the world. During grasping, the object blocks cameras, fingers cover each other, and the moment that matters is hidden from an external sensor.

That is why end-effector sensing changes the evaluation frame.

Ommo Technologies is a useful example. According to the report, the company chose a permanent-magnet magnetic-field positioning route after comparing optical positioning and millimeter-wave radar. Zheng Minjie told the outlet that mainstream optical positioning fails when objects block the line of sight. Ommo's route uses mechanical rotation of a permanent magnet to generate a characteristic magnetic field, then calculates complete 6DoF position and posture data from real-time magnetic-field signals collected by miniature magnetic sensors.

The important detail is size. Ommo's smallest magnetic sensor is about 0.8 millimeters, and the company says its sensors can be embedded in robot end-effectors, wearable devices, or medical devices. Its system can currently reach sub-millimeter accuracy within a specified working range. That turns the hand from a moving tool into a data instrument, especially when the same sensing approach appears in Ommo's first data collection glove, which uses medical-grade permanent-magnet sensors to collect human hand operational posture data.

CASBOT's dexterous-hand lineup shows the other side of the shift. The same report describes D1 as a high-degree-of-freedom general-purpose five-finger dexterous hand, with emphasis on motion flexibility, independent multi-joint movement, multi-finger coordination, and tactile sensing. M1 is built as a modular scenario hand, using a standardized main body plus functional fingertips for rapid switching. The F series explores high degrees of freedom and compliant manipulation through a tendon-driven biomimetic structure. Those designs matter most when their motions can be captured as reliable contact trajectories, not just demonstrated on video.

The data bottleneck has two competing starting points

The bottleneck looks different depending on where a team stands in the data loop. Ommo Technologies starts with the moment of capture. Zheng Minjie told the outlet that mainstream optical positioning fails when objects block the line of sight, which is exactly the kind of failure a hand creates for itself when fingers, tools, and parts crowd the workspace.

Ommo's answer is to move tracking into the object-contact zone. According to the outlet, its system calculates complete 6DoF position and posture data from magnetic-field signals collected in real time by miniature magnetic sensors, and can currently reach sub-millimeter accuracy within a specified working range. The company says those sensors can be embedded in robot end-effectors, wearable devices, or medical devices. Its first data collection glove uses medical-grade permanent-magnet sensors to collect human hand operational posture data.

That is one starting point: if tracking breaks, the dataset is compromised before software can clean it.

Qiongche Intelligence's starting point is later in the chain. According to geekpark, Qiongche uses the UMI approach for data collection, with a device that synchronously records first-person-view images, gripper poses, and motion trajectories. But geekpark also describes the raw capture as only the beginning. Before the data can enter model training, it must go through pose recovery, fisheye camera calibration, action segmentation, and multimodal annotation.

For Qiongche, the choke point is not only seeing the manipulation event. It is processing the event at cloud scale. Geekpark says Visual SLAM reconstruction consumes large amounts of GPU computing power in Qiongche's workflow, so Qiongche and Alibaba Cloud built an end-to-end UMI Data + AI Pipeline. Alibaba Cloud MaxCompute MaxFrame splits video undistortion, SLAM pose recovery, and multimodal annotation across cloud resources for parallel execution.

The contradiction is useful. Ommo says poor capture poisons the source; Qiongche says unprocessed capture stalls before training. Geekpark reports that Qiongche achieved full-process automation, more than 10 times higher overall data processing throughput, and elastic scaling to more than 100,000 CU through Alibaba Cloud.

A useful manipulation dataset needs a workflow, not a folder of videos

A useful manipulation dataset starts before the model sees anything. According to geekpark, Qiongche Intelligence frames its robot infrastructure as data-training-deployment, and its UMI collection method begins with an operator holding a device while doing ordinary manipulation: grasping, opening doors, carrying objects. That matters because the capture format is already tied to the end effector. UMI synchronously records first-person-view images, gripper poses, and motion trajectories, so the dataset is not merely video evidence of a task. It is a record of what the hand saw and how it moved.

The folder is the wrong mental model.

Geekpark describes the next layer as a processing chain that robot teams should audit item by item. Raw operation data has to pass through pose recovery, fisheye camera calibration, action segmentation, and multimodal annotation before training. Qiongche Intelligence and Alibaba Cloud built an end-to-end UMI Data + AI Pipeline for that work, with MaxCompute MaxFrame splitting video undistortion, SLAM pose recovery, and multimodal annotation across cloud resources for parallel execution. Large-model capabilities are embedded in the workflow to semantically annotate action clips automatically.

The less glamorous pieces are what make the loop repeatable. DataWorks orchestrates the whole processing workflow, including periodic scheduling, historical data reruns, and automatic exception recovery on a unified platform. Processed data then enters Hologres for unified management, where it supports real-time sample retrieval, data review, and training management. After that, the data enters PAI for model training, performance optimization, and deployment validation.

That schema turns collection into a production system: capture, clean, label, retrieve, review, train, validate. It also explains why the outlet's note on CASBOT matters. CASBOT proposed a dual-flywheel logic of multi-series bodies x data collection, and said it would help build a specialized dexterous hand skills training ground for data collection, skills training, and application verification around real tasks. The value is not another clip of a hand succeeding once. It is the ability to reproduce the path from task attempt to usable skill evidence.

Task validation is the antidote to parameter competition

CASBOT is a useful case because its story does not stop at another hand specification sheet. According to media reports, the company debuted in Changsha, Hunan on August 7 and released three dexterous hand products, but its stated direction is to turn dexterous hands from single hardware components into a software-hardware manipulation capability platform.

That shift matters more than the product spread. The L1 takes a lightweight route with simplified degrees of freedom. The D1 is positioned as a high-degree-of-freedom general-purpose five-finger hand. The M1 uses a standardized main body plus functional fingertips, so different operating capabilities can be switched in and out.

Specs still have a role. They do not prove skill.

The more important row is CASBOT's proposed dual-flywheel logic of "multi-series bodies x data collection." Media reports also say that the company will participate in building a specialized dexterous hand skills training ground for data collection, skills training, and application verification around real tasks. That is the right unit of progress: a task where the hand drops the object, slips on a surface, misses a grasp, recovers, or fails to recover.

A robot band can show range. Media reports say CASBOT's L-series hands have been used by the CASBOT BAND robot band to play guitar, bass, and electronic keyboard. But the manufacturing lesson is not that musical demos settle the question. It is that repeated contact gives teams a way to observe timing errors, finger placement, and recovery behavior under a defined task.

This is where parameter competition starts to look thin. GGII's report, cited in media coverage, put China's dexterous hand market sales at about 19,200 units in 2025 and forecast 70,200 units in 2026. More hands will create more trials, but only if teams capture the failures as carefully as the successes.

Choose tools by where your manipulation-data loop is weakest

• Your data bottleneck is accurate hand or end-effector pose capture, especially where cameras lose line of sight. Consider an end-effector sensing route like Ommo Technologies' permanent-magnet positioning system: the sources say it calculates complete 6DoF position and posture data from real-time magnetic-field signals, currently reaches sub-millimeter accuracy within a specified working range, and uses magnetic sensors as small as about 0.8 millimeters. This is most relevant when optical positioning is blocked by objects; it is less justified if your manipulation tasks can already be captured reliably with cameras.
• Your team can collect demonstrations but cannot turn raw video and trajectory data into training-ready datasets fast enough. Prioritize an automated data pipeline like Qiongche Intelligence's UMI Data + AI Pipeline with Alibaba Cloud: the ledger says raw operation data needs pose recovery, fisheye camera calibration, action segmentation, and multimodal annotation before training, and that MaxCompute MaxFrame parallelizes video undistortion, SLAM pose recovery, and multimodal annotation. This is a data-engineering investment, not just a robot-hardware purchase.
• Your manipulation roadmap requires validating specific skills, not just buying a dexterous hand as a component. Use a hand-plus-validation approach like CASBOT embodied intelligence's stated plan: it proposed a dual-flywheel logic of "multi-series bodies x data collection" and announced participation in building a specialized dexterous hand skills training ground for data collection, skills training, and application verification around real tasks. This fits teams that need repeatable task validation; it is less suitable if you only need a standalone gripper.
• You need different hand designs for different manipulation regimes. Map the hand to the task class rather than choosing one universal option: CASBOT's L1 uses a lightweight route with simplified degrees of freedom; D1 is a high-degree-of-freedom general-purpose five-finger dexterous hand; M1 uses a standardized main body plus functional fingertips for rapid switching; and the F series explores tendon-driven biomimetic, high-degree-of-freedom compliant manipulation. The sources do not disclose performance benchmarks or prices for comparing them.
• Your goal is consumer engagement, creator customization, or education rather than industrial manipulation-data collection. AgiBot Q1 Exploration Edition fits that lane better than a production manipulation platform: the sources describe open-source exterior structural parts, full-body 3D-printed customization, modular replacement, and a zero-code visual platform for appearance modeling and motion choreography. Do not treat the Q1 evidence as proof of an industrial manipulation-data loop; the ledger frames it mainly around personal robots, customization, and technology designer toys.

Consumer robots may win attention before they win tasks

AgiBot's Q1 Exploration Edition shows why a consumer robot can attract attention before it proves much task value. At ChinaJoy, ifanr reported shells themed as a Space Marine, an armored fighter, a superpowered mecha, and an astronaut. The largest crowd gathered around a World of Warcraft murloc version, which also visited the Blizzard booth beside a human murloc cosplayer.

That is not a manipulation benchmark. It is a market signal.

The Q1 Exploration Edition is AgiBot's first-generation personal prototype robot, standing 88 centimeters tall and weighing about 15 kilograms, according to ifanr. Its pitch is emotional participation: fully open-source exterior structural parts, full-body 3D-printed customization, modular replacement, and a zero-code visual platform for appearance modeling and motion choreography. Users can change the shell, movements, personality, voice, expressions, and behavior logic.

ifanr describes Q1 as the world's first personal robot with DIY character customization, and also as the world's first small-size humanoid robot with full-body force control. Its self-developed miniature QDD quasi-direct-drive joints are meant to preserve force-control performance and high dynamic response from larger models. During the 4-day ChinaJoy exhibition, tens of thousands of players visited AgiBot's booth, which supports ifanr's argument that technology designer toys may be a realistic route for personal robots into the consumer market.

Customization can create attachment. ifanr cites a 2024 human-computer interaction experiment in which participants involved in customization showed stronger psychological ownership of robots. It also cites Bambu Lab data showing that 90% of users were still printing 12 months after buying a 3D printer. AgiBot is aiming beyond companies with engineering teams, toward independent creators, designer toy enthusiasts, and students.

CASBOT is walking a different path. A Chinese outlet reports that its embodied intelligence work starts from dexterous hands and manipulation capability, with a goal of turning hands from single hardware components into a software-and-hardware capability platform. Its stated development logic is "multi-series bodies x data collection," and it plans to help build a specialized dexterous-hand skills training ground for data collection, skills training, and application verification around real tasks.

These paths can coexist. Character sells the body; validated manipulation makes it useful. The deeper shift is that the winning robot project will treat the hand, the training ground, and the data pipeline as one product system.

Embodied-AI workflow pieces across sensing, capture, processing, and validation

Watch for the teams that treat the hand sale as the start of the loop, not the finish. The outlet reports GGII's forecast of China dexterous hand sales moving from about 19,200 units in 2025 to 70,200 units in 2026. That growth will make demos easier to stage and harder to trust.

In your own roadmap, ask for the missing venue and the missing data path. CASBOT's F series points to tendon-driven, biomimetic, high-degree-of-freedom compliant manipulation, but its more useful clue is organizational: a dual-flywheel of multi-series bodies and data collection, plus a specialized skills training ground for real-task collection, training, and verification. Treat insertion, rotation, tool use, and deformable-object handling as acceptance tests.

For readers outside China

• Availability: Availability outside China is not disclosed for most products. Ommo Technologies' cooperative customers currently cover more than 100 domestic and overseas medical device companies, but product availability, export access, and sales channels are not disclosed in sources. Qiongche Intelligence is described as a Chinese embodied intelligence company working with Alibaba Cloud. CASBOT embodied intelligence has settled in Hunan Xiangjiang New Area. AgiBot Q1 Exploration Edition appeared at ChinaJoy, but overseas availability is not disclosed.
• Pricing: Product pricing is not disclosed in sources. The only financing figure given is that Ommo Technologies completed a Series A financing round worth tens of millions of dollars. No prices are disclosed for Ommo's sensing system or glove, Qiongche Intelligence's pipeline, Alibaba Cloud usage, CASBOT's dexterous hands, or AgiBot Q1 Exploration Edition.
• Closest Western equivalents: Not disclosed in sources for Ommo Technologies' magnetic positioning system; UMI, or Universal Manipulation Interface, for Qiongche Intelligence's data collection approach; Not disclosed in sources for CASBOT's dexterous hand product matrix; Not disclosed in sources for AgiBot Q1 Exploration Edition
• Data residency: The Qiongche Intelligence workflow uses Alibaba Cloud services including MaxCompute MaxFrame, DataWorks, Hologres, and PAI. Processed data enters Hologres for unified management and then PAI for model training, performance optimization, and deployment validation. The sources do not disclose cloud region, cross-border data transfer arrangements, retention policies, or whether data remains inside China.

Sources

• 36kr 赋予机器人"空间直觉"，「Ommo Technologies」获数千万美元A轮融资｜36氪首发 https://36kr.com/p/3927419946629256
• geekpark 打破机器人「数据焦虑」，穹彻智能想建一条「训练数据」生产线 https://geekpark.net/news/368553
• ifanr 在 ChinaJoy ，我看到了第一台「科技潮玩」机器人 https://ifanr.com/1674308
• 36kr 最前线|中科慧思发布三款灵巧手产品，要从"抓取"走向真实场景作业 https://36kr.com/p/3931010805579139