IROS 2026 Observation: How Yuansheng Intelligence's Apex Hand Takes Dexterous Hands from Demo to Productivity Tool

The design logic of a dexterous hand must be reverse-engineered from real-world deployment needs.

English text is being prepared.

English text is being prepared.

At this year's IROS, dexterous manipulation and tactile sensing were hotter than ever. Beading, USB insertion, hot-stamping a logo onto a canvas tote bag, in-hand pen spinning, ball spinning, and cocktail mixing—various demos played out in succession on site. From ICRA to WRC to IROS, we have tracked nearly all the top robotics conferences this year, and one impression has become increasingly clear: for dexterous hands, completing complex motions is increasingly no longer difficult. Demonstrations are no longer limited to simple grasping and transport; they increasingly emphasize contact-rich manipulation, dynamic adjustment, long-horizon operation, and deformable object handling—tasks closer to real work. Overall, manufacturers' demand for deployment is growing, and their mindset is more pragmatic. In the past, everyone pursued rich parameters and extreme metrics; now many parameters are starting to be dialed back—some have reduced degrees of freedom, and some have made sensor configurations more pragmatic. Long-term reliability, stable use, ease of maintenance, and whether the product can truly be put to use after delivery have been given higher priority. This is also an interesting change we saw at the Yuansheng Intelligence booth this year. At the IROS site, Yuansheng Intelligence's high-DOF five-fingered dexterous hand, Apex Hand, demonstrated a set of manipulation tasks with different emphases: reinforcement-learning-based in-hand ball spinning, pulse-taking with a dexterous hand, real-time full-palm tactile sensing, continuous high-load constant-temperature grasping, and high-impact operations such as splitting wooden boards and hammering nails. These demos showcased Apex Hand's comprehensive capabilities in contact-rich manipulation, dynamic adjustment, long-horizon operation, and high-impact scenarios. It can already complete sufficiently complex motions, and in some dimensions even goes beyond the scope of conventional grasping and transport. But how can these capabilities move from the booth to real work and be continuously used by customers? What comprehensive capabilities does a dexterous hand aimed at real applications actually need? Out of curiosity, we spoke with Yang Sicheng, founder of Yuansheng Intelligence, and found that Yuansheng's thinking on products and deployment is very clear: from the very beginning, Yuansheng has defined its products from the perspective of deployment, starting from the needs of real scenarios and reverse-engineering what capabilities a dexterous hand needs. From product definition to mass-production validation, and then to data collection and the development ecosystem, his thinking on the productization of dexterous hands can be summarized in three judgments: First, competition in dexterous hands is shifting from pursuing extreme performance metrics to building system-level capabilities for real tasks. So-called system capability does not mean the higher a single metric the better; rather, it means achieving balance across dimensions such as reliability, dexterity, load capacity, dynamic response, tactile sensing, and precision, with no obvious weak points. Second, once parameters are sufficient, the real dividing line is the ability to work stably over the long term. Heat dissipation, structural reliability, fatigue life, and maintainability—problems that are not easily seen in demos—will ultimately affect maintenance costs, lifespan, and downtime risk. Third, hardware delivery does not mean delivery of manipulation capability. Without solving the problem of acquiring high-quality manipulation data, "no matter how good the hardware is, it will be hard to use." These three judgments also provide a lens for observing changes in the dexterous hand industry this year.

01

What capabilities must a deployable dexterous hand actually have? The first hurdle for deploying a dexterous hand is product capability itself. Over the past few years, dexterous hand products have continuously pushed the boundaries of metrics such as degrees of freedom, payload, and precision. But as more and more manufacturers are able to demonstrate complex five-finger manipulation, a question has become increasingly prominent: does an improvement in parameters necessarily mean an improvement in actual manipulation capability? Yang Sicheng has his own judgment on this. In his view, for a dexterous hand aimed at real applications, the first step is to clarify which capabilities truly affect task completion, and which performance improvements bring additional structural complexity and engineering cost. Only on that basis do parameter improvements have meaning for actual manipulation capability. Looking back at the product definition of the Apex Hand, Yuansheng made several ahead-of-their-time choices: the trade-off in degrees of freedom, the minimum grasp diameter, and full-palm tactile sensing and all-joint backdrivability. The designs behind some of these metrics were not mainstream at the time, and only now are they gradually becoming industry consensus. The first choice is controlling degrees of freedom. The Apex Hand's 21 degrees of freedom correspond to the anatomical structure of the human hand: 5 for the thumb and 4 for each of the other four fingers. It is also one of the very few mass-produced five-finger dexterous hands in the world that can pass the Kapandji opposition test with a perfect score. Many in the industry once built 22 degrees of freedom, with the extra one being the hypothenar opposition degree of freedom near the little finger. Many manufacturers used to include this degree of freedom, but this year they have chosen to remove it. It can increase the local motion capability of the palm, but for most practical manipulation tasks, the performance gain it brings is relatively limited while requiring additional mechanical structure cost. In Yuansheng's definition, more degrees of freedom is not necessarily better; the key is to retain the core degrees of freedom that are truly useful for manipulation. The second choice is to incorporate minimum grasp diameter into the product definition. In addition to degrees of freedom, Yang Sicheng mentioned that a common customer feedback is that the Apex Hand looks slightly larger overall than a human hand. "Large hands" easily raise concerns that fine manipulation will be limited, but in reality, more important than hand size is the minimum grasp diameter—an often overlooked parameter that is the key to fine manipulation. The minimum grasp diameter determines how closely a hand can bring its fingers together, and also determines the lower limit of object sizes it can manipulate stably. The smaller the minimum grasp diameter, the broader the range of objects a dexterous hand can cover. If the minimum grasp diameter is too large, the fingers cannot close fully into position and can only barely pinch, resulting in unstable contact and easy slipping, which limits fine manipulation and in-hand adjustment. Many manufacturers only began to pay attention to this point this year, whereas Yuansheng designed its product from the start with a minimum grasp diameter of about 5 millimeters. At the exhibition, the Apex Hand was able to grasp a slender pen, take a business card from a card holder, pick up a phone from a tabletop, and complete in-hand manipulation. The third choice is to consider full-palm tactile sensing and all-joint backdrivability capabilities early in product design. For real manipulation, the robot also needs to know what is happening between its fingers and the object. The value of tactile sensing is not just adding another perception dimension, but providing the basis for the robot to continuously adjust its actions according to contact conditions. This is also why tactile perception and joint backdrivability are attracting increasing attention. Both are features that many manufacturers only added to their products this year. Yuansheng, by contrast, implemented full-palm tactile coverage early on, while all-joint backdrivability allows the fingers to flexibly adjust their grasp state according to external loads and maintain stable contact under complex force conditions. At the IROS site, Yuansheng demonstrated "pulse diagnosis with a dexterous hand," turning tactile perception capability into a more intuitive interactive experience. These understandings of dexterity, fine manipulation, and perception reflect Yuansheng's overall judgment about dexterous hand products: tasks in real scenarios are hard to simply classify as "moving heavy objects" or "performing fine manipulation," and a general-purpose dexterous hand needs to handle tasks at different scales and with different force requirements simultaneously. Yang Sicheng summarizes this product approach as a balanced pursuit of six dimensions: reliability, dexterity, payload capacity, dynamic response, tactile perception, and precision. For a dexterous hand aimed at practical applications, an obvious shortcoming in any one dimension may limit overall task capability. By his standard, every item needs to reach 80 or 90 points or above for the hand to become a "relatively general-purpose, universal productivity tool."

02

Once the specifications are sufficient, how can long-term operation be achieved? Once product capability passes muster, the next question is: can this hand work continuously? In the lab, completing a complex operation once is already enough to prove a certain technical capability. But once it enters real tasks, robots often need to repeatedly perform operations over long periods, continuously bear loads, and cope with unexpected contact and impacts. This is why Yuansheng puts "reliability" first in its product definition. Yang Sicheng mentioned that many dexterous hands still emphasize what movements they can complete, but for real deployment, reliability, complex operation capability, long-term use precision, and later maintainability will be more important. Demonstrations at industry exhibitions tend to favor instantaneous movements and peak capability; displays that can run continuously for several hours and remain stable under load are rare. At the IROS site, on the Yuansheng booth we saw a demonstration that looked simple but that we had never seen at any other booth: the Apex Hand maintained an active grasp on two large bottles of mineral water all day, weighing more than 10 kg, from after the exhibition opened in the morning until the evening, 8 hours without interruption, with temperature basically remaining stable throughout. Among products of the same class, we have hardly seen any dexterous hand that can continuously grasp under the same load for more than 10 minutes. What this demonstration shows is the reliability of a dexterous hand under sustained load: it must have enough strength to keep grasping continuously, and also control the temperature rise caused by long-term operation to ensure stable output. Yuansheng believes that the load of a dexterous hand is not only about carrying weight, but also includes grip force. Unscrewing a bottle cap, tearing packaging, holding a knife to cut hard objects—these task objects are not heavy, but all require the robot to actively apply force; without enough grip force, the action cannot be completed; more importantly, whether it can continuously output force while using tools. Yang Sicheng adapted a saying to summarize: "If a hand cannot even truss a chicken, a dexterous hand cannot become productivity." The Apex Hand's maximum single-hand active load is 30 kilograms. At the second World Humanoid Robot Games in August this year, the gold, silver, and bronze medal teams in the weightlifting and tug-of-war heavyweight categories were all equipped with the Apex Hand; on the matter of "active load," this hand has already undergone a round of real-world testing. Sufficient strength is only the foundation; whether it can stably output under sustained load and control temperature rise is the key to the reliability of a dexterous hand. Applying force for a long time increases the burden on the drive system and transmission mechanism, and temperature changes may also affect motion precision and tactile perception. Being able to control temperature rise and maintain stable output under sustained load is closer to the requirements of practical applications. For a dexterous hand that integrates motors, transmission mechanisms, and tactile sensors, thermal management involves multiple links. Excessive temperature may trigger overheating protection, causing equipment shutdown, and in severe cases may damage motors and accelerate aging of insulation materials. Temperature changes may also cause drift in tactile sensors, reducing measurement precision; long-term accumulated problems such as thermal deformation may affect motion precision and service life. Yuansheng chose to solve this problem starting from the drive scheme. In terms of drive schemes, common industry routes each have trade-offs. Traditional direct-drive schemes usually distribute motors to each finger, making heat sources more evenly distributed and heat dissipation seemingly easier. Yuansheng instead chose a hybrid drive scheme based mainly on tendon drive (tendon-sheath transmission), combining direct drive and linkage transmission. The motors are concentrated inside the palm, and the fingers are pulled by tendons; this design keeps heat sources away from fingertip tactile sensors; on the other hand, once heat sources are concentrated, centralized heat dissipation is actually easier to handle. Beyond reliability, control difficulty and maintenance cost also determine whether a dexterous hand can be put into long-term use. When tendon drive is mentioned, many people still think that tendon-driven hands are harder to control than direct-drive hands; Yang Sicheng considers this a misunderstanding. The reason for this misunderstanding is that early Shadow's tendon drive scheme was indeed troublesome, with problems such as nonlinear friction and creep, and required regular calibration. Yuansheng's hybrid drive, based mainly on tendon drive, is already very different from early tendon drive schemes; the actual effect is closer to direct drive, avoids complex transmission, and has joint force transparency superior to most direct-drive hands, making control friendlier instead. In terms of maintenance, its structure considers modular design, so a single unit can be extracted and replaced without disassembling the entire finger. These considerations will translate into differences in maintenance efficiency, operational stability, and usage cost after long-term product operation. Yuansheng adopts an intra-palm hybrid parallel scheme; the tendon is only one finger long, so it can be individually extracted and replaced without disassembling the finger. Instantaneous impacts in real tasks are also challenges that dexterous hands need to face. This is also why, in many industrial scenarios, grippers with relatively simple structures and higher reliability are still widely used. Compared with five-fingered dexterous hands, grippers can complete a limited range of operation types, but in highly repetitive tasks, their structure is simple, maintenance is convenient, and they can withstand certain collisions. At the IROS site, Yuansheng demonstrated high-impact tasks such as hammering nails and splitting wooden boards; after completing high-impact operations, the Apex Hand's hand structure remained intact, and its movement remained smooth and precise. Of course, to prove that a dexterous hand can work stably over the long term, product reliability needs to be gradually established through extensive repeatable engineering verification. Yang Sicheng introduced that in the process of advancing Apex Hand toward mass production, Yuansheng conducted extensive testing and optimization around structural life, transmission reliability, and continuous working capability. One important task was million-cycle loaded fatigue testing. A single complete test requires months of continuous operation. During testing, the team also needs to locate problems based on test results, optimize structural design, transmission systems, or related parameters, and then carry out verification again.

03

After delivery, how do customers actually put it to use? The third hurdle lies beyond the booth: once a hand has been delivered, how do customers get the robot to actually do work? Hardware having manipulation capability does not mean the robot has already mastered those operations. From grasping objects of different shapes to using tools to complete complex tasks, robots still need data and training to acquire adaptability to different manipulation scenarios. At this year's IROS, one very obvious change is a marked increase in data collection and teleoperation solutions built around dexterous hands. What this reflects is another threshold the industry is facing: as dexterous hand hardware gradually matures, how can real manipulation data usable for model training be obtained efficiently? Dexterous manipulation places very high demands on data quality; especially when finger coordination, contact, and fine manipulation are involved, the degrees of freedom and mapping relationships of the collection equipment all affect whether the final data can be used for training. Yang Sicheng particularly emphasized the importance of data quality: making data serve actual tasks is more important than piling up data volume. In his view, some solutions (for example, using low-DOF dexterous hands to collect data) may reduce upfront collection costs, but if such data ultimately cannot be mapped onto the target dexterous hand, it will still be difficult to support training of complex manipulation models, and the earlier investment may lose its meaning. For Yuansheng, this is also a problem that must be continuously solved after the product enters practical application. Yang Sicheng introduced that the team is advancing a second-generation data collection solution and is cooperating with third-party data collection companies and data glove manufacturers, with the goal of lowering the threshold for customers to use dexterous hands for task development through coordination among hardware, teleoperation, and data collection solutions. Lowering the thresholds for hardware acquisition and development is also an important condition for the further development of dexterous hand technology. If research teams must first solve problems of hardware acquisition, system integration, and basic development, the resources left for algorithm research and task validation will be limited. During IROS, Yuansheng Intelligence officially released the "Yuansheng Origin Program" for universities, research institutions, and developer teams worldwide, opening free loan of Apex Hand prototypes and providing development kits and dedicated technical support. For the dexterous manipulation industry, which is still in a stage of rapid development, how to get more developers involved in exploration and gradually build a development ecosystem around the hardware platform may be another important task in moving dexterous hands from a single product toward broader application. In terms of specific application directions, the commercialization of dexterous hands still needs to advance gradually according to task difficulty and technological maturity; sorting and assembly are the scenarios Yuansheng is currently focusing on exploring. Yang Sicheng believes that sorting has requirements for model generalization capability, but not excessively high ones, and can also first decouple upper-limb manipulation from lower-limb movement, making it a direction that can be quickly brought to work at present, while assembly, which demands stronger part manipulation capability, is more difficult, and truly stable deployment may still require extensive technical iteration. From the initial product definition around degrees of freedom, payload, touch, and precision, to advancing mass production this year, and verifying product reliability through continuous load, high-impact operation, and million-cycle loaded fatigue tests, Apex Hand is moving from capability verification toward more systematic engineering verification. Returning to the question at the beginning, for dexterous hands to reach deployment, at least three thresholds must be crossed: whether manipulation capability is sufficient, whether they can work stably over the long term, and whether customers can enable the robot to continuously learn new tasks. Impressive demos are increasingly close to the starting point; what truly determines how far a hand can go is whether it can complete delivery, enter specific tasks, and continue working in real environments. For Apex Hand, this verification has already begun. Leiphone

To make it easier for everyone to team up for communication and exchange conference news, we have specially created an IROS 2026 attendee exchange group! Joining the group unlocks these "benefits"? (Leiphone)

Conference Intelligence Station: submission DDLs, format specifications, review progress... key milestones delivered as soon as they arrive, so you never miss a deadline and can prepare your submission with confidence.

Peer Tea Party: peers from all over are in the group, chatting about experiences, exchanging ideas, and building connections; on the research path, we are with you.

Frontier Supply Station: the latest research results and hot directions synchronized in real time, helping you sort out your submission thread in light of the conference theme and fueling your paper inspiration.

On-site Support Team: (exclusively opened during the conference): no need to panic once you arrive at the venue—

Topic Co-reading: popular sessions and Keynote discussions, so you can catch up even if you miss them;

Poster Compilation: highlights of on-site posters organized for you, one-click collection so nothing is missed;

Meetup Plaza: dinner meetups, interview meetups, exchange meetups... whether you want to chat, collaborate, or meet in person, just initiate it in the group.

? Group entry portal: scan the QR code to join the group or add WeChat Luyoyo_2026, with the note: IROS + institution/school + name + direction.