IROS 2026 in Pittsburgh Made the Case That Robot Hands, Not Legs, Are the Bottleneck

Robotics & Drones 3 OCT 2026 2,001 words 10 min read By news.jualin.id

IROS 2026 in Pittsburgh Made the Case That Robot Hands, Not Legs, Are the Bottleneck

Introduction

For most of the last decade the headline claim in humanoid robotics has been about locomotion. Vendors have competed on how a machine walks, balances, climbs stairs and survives a shove, and a long parade of demonstration videos has treated bipedal competence as the gating problem. The IEEE/RSJ International Conference on Intelligent Robots and Systems, which ran from September 27 to October 1, 2026 at Pittsburgh's David L. Lawrence Convention Center, made the case that the centre of gravity has moved. The dominant story on the exhibition floor was not walking. It was the hand.

That is not a cosmetic change of subject. Reaching an object and picking it up are different problems, and only one of them has been substantially solved. A robot that can cross a warehouse floor reliably but cannot seat a connector, thread a fastener or recover a grip when a part slips is a mobile camera on legs. At IROS 2026, essentially every major manipulation announcement circled the same question: what does it take to move an object through a hand, under contact, repeatedly, without a human intervening?

The conference programme itself gave the trend its scale. The official IROS 2026 programme carried more than 1,900 contributed papers, 19 keynote talks, 86 workshops and tutorials, and 9 competitions, hosted in a city the organisers nickname "Roboburgh" under general chair Howie Choset of Carnegie Mellon University. A week built around 1,900 papers would be expected to bury any single theme. This year, the trade floor made the theme legible: dexterity has become the product category.

The Four-Finger Decision

The loudest single piece of news to land around the conference came from Boston Dynamics, which unveiled new hands for its Atlas humanoid. The mechanical design is more interesting than the degree-of-freedom count, because the interesting decision was what the company left out.

Atlas previously carried seven-degree-of-freedom hands built to grasp a wide variety of objects. The new generation, identified in the company's technical material as GR3, moves to 13 degrees of freedom across four fingers including a more capable opposable thumb. The thumb carries four degrees of freedom and each of the other three fingers carries three, allowing the fingers to splay and the thumb to slide along and across them for pinches, three-point grasps and triggered tool grips. Robotics 24/7's report on the launch lists the intended behaviours: sliding the thumb fingertip along the length and width of the other fingers, dexterous pinch and tripodal grasps, and powered tool grasps including drills, torque drivers, grinders, nail guns and welding torches.

The fifth finger was considered and rejected. According to the technical write-up, the team taped their own ring and little fingers together to assess the practical loss, and concluded that the additional three actuators were not worth the cost, the volume, or the extra failure opportunities. The joints use a single actuator type, fully encapsulated, with no fragile cables crossing them, and the whole assembly is roughly the size of a large human hand so that it fits human-scale tools and workspaces. Boston Dynamics says strength is comparable to the previous hand, referencing Atlas carrying a loaded mini-fridge weighing more than 100 pounds.

Two engineering arguments underneath that choice matter more than the DOF number. First, simulation fidelity: the rigid, backdrivable drive is meant to be easy to model, so friction, motor characteristics, object shapes and disturbances can be varied in simulation before a policy transfers to hardware. Alberto Rodriguez, Atlas's director of AI and robot behaviour, framed reinforcement learning in simulation as an essential complement to imitation, because fast manipulation needs rapid feedback and force regulation that human demonstrations alone do not supply. Second, backdrivability as a durability feature: a joint that yields under an external impact can protect its own gearbox. That is a design rationale rather than a published impact rating, and Boston Dynamics reports promising initial transfer results without providing a broad task-success benchmark.

Hardware and Its Teaching Interface

A hand cannot be trained well by a hand that does not exist. That constraint produced the most coherent pattern at IROS 2026: companies shipping manipulation hardware alongside the interface used to teach it, so that the shape of the demonstration matches the shape of the deployment machine.

Sharpa announced three products at the conference under the banner of dexterous manipulation. D01 is an integrated robot built around Sharpa Wave dexterous hands, seven-degree-of-freedom arms and whole-body electronic skin, specifying an approximately 1:1 arm payload-to-weight ratio, end-effector speed above 10.5 metres per second, 1,000 Hz communication frequency and 0.2 mm end-effector repeatability. W02 is a fully tactile dexterous hand with 21 active degrees of freedom under 750 grams, pairing fingertip sensors with a 5 millinewton to 30 newton range and 1 mm spatial resolution against electronic skin across the fingers and palm at 0.1 to 20 newtons and 5 mm resolution. AE01 is the piece that explains the rest: a haptic exoskeleton data glove with 22 degrees of freedom, 22 encoders, 256 levels of vibrotactile feedback per fingertip, and fingertip position repeatability under 1 mm. The company states it requires no additional user calibration.

Chestnut Robotics took the same shape with a different emphasis, launching the Aero Hand alongside Aero UMI, a wearable capture and teleoperation rig. The hand is human-hand-sized, combining tendon, direct and linkage actuation with a claimed lifetime above three million cycles; the capture rig uses magnetic encoders and a matching URDF description of the robot, with 0.1-degree joint measurement accuracy. Co-founder and CEO Evan Tao announced the pair in a September 28 post, explicitly placing the work in the Universal Manipulation Interface lineage — the research approach of collecting demonstrations on handheld devices away from the robot and training policies that then run on the arm. Chestnut describes the pairing as having "zero embodiment gap," which is a design claim about matching morphology and tactile sensing, not a published demonstration that demonstrations transfer without error.

The rationale for building the collector around the deployment hand is straightforward once stated. Coordinating several fingers around an object contains far more choices than closing a parallel-jaw gripper, so a human demonstration can contain motions and contacts the target robot cannot reproduce. Sharpa's assertion that robots "should assist, not replace humans" is a mission statement; Chestnut's is an engineering one. Both arrived at the same conclusion from different directions.

Fewer Fingers, More Sensors

Not everyone concluded that more fingers is the answer. DH-Robotics showed a portfolio organised by task rather than by anatomy, including the ADH-5-13, a five-finger hand with 13 independently controllable active degrees of freedom. The thumb, index and middle fingers carry three degrees of freedom each, the ring and little fingers two each, and the index and middle add plus or minus 15 degrees of lateral motion for multi-finger poses. At 735 grams with a maximum lifting payload of 20 kg once an object is grasped, the hand adds fingertip tactile sensing, replaceable modules that can be swapped without dismantling the whole hand or a full finger, and a claimed 300 percent improvement in maintenance efficiency. Its companion UDH-3-7 is a three-finger, seven-degree-of-freedom unit rated to 15 kg with 50 newtons of fingertip force and plus or minus 0.02 mm positioning repeatability over a one-million-cycle design life.

The most interesting dissent came from a Columbia University spinout that announced its funding during the conference week. Tangent Robotics raised $4.5 million in pre-seed, led by Fly Ventures and Toyota Ventures with participation from Logos Fund and Sparked Ventures, to build fine motor skills for manufacturing tasks such as fitting connectors, seating gaskets, threading and assembling gears. Founded by Pedro Piacenza, Matei Ciocarlie and Ioannis Kymissis out of Columbia Engineering's Robotic Manipulation and Mobility Lab, the company positions itself as a fine motor skill layer sitting between vision-language-action models and low-level controllers. Its TR02 hand, introduced in June, has three fingers and ten degrees of freedom with optical touch sensors at each fingertip and an offset palm camera for visual feedback. The third finger and palm exist, the company says, so the hand can hold a jar while opening its lid or move an object from a fingertip grasp into the palm.

Two smaller signals from the floor point the same way. Boston Dynamics' own researchers have been publicly mapping the trade-offs in tendon-driven hands, from wrist friction and wear to control at contact. And the technical programme devoted a dedicated block of industry talks to tactile sensing infrastructure — piezoresistive sensors, high-frequency vision-based tactile sensing, tactile-native world models — from vendors such as Daimon Robotics, Apex Sensing, Tashan Technology and DexRobot. Tactile data collection was treated as infrastructure, not as a feature.

The Volume Question

None of this settles whether dexterity is solved at scale. The clearest available volume data cuts in an interesting direction. IDC figures reported through Robotics & Automation News put Agibot at more than 8,600 humanoid units shipped in the first half of 2026, roughly 35 percent of a global market approaching 25,000 units, with worldwide humanoid shipments up 432.1 percent year on year. Agibot separately said it reached 20,000 cumulative robots produced in September 2026, and disclosed more than 300 robots deployed at Chimelong Spaceship Park in Hengqin plus 100 X2 humanoids across ASD stores.

Those numbers establish manufacturing and sales scale more clearly than they establish technical performance. The published figures do not come with a full vendor ranking, a stated methodology, or a definition of what qualifies as a humanoid shipment, and the deployments come with no measures of autonomy, intervention rates, uptime or task completion. A robot greeting customers in a store is a real deployment; it is not a demonstration that the hand can hold a part through a thousand cycles without a technician.

Conclusion

The through-line of IROS 2026 is that the industry's hardest unsolved problem is contact, and that vendors have started treating it as a system problem rather than a component problem. Boston Dynamics is spending degrees of freedom to buy manufacturability and simulation fidelity. DH-Robotics is making the hand serviceable so downtime is a scheduled event rather than an incident. Tangent is abandoning human finger count entirely on the argument that tasks determine hand architecture. Sharpa and Chestnut are shipping the data-collection rig in the same box as the hand, because demonstrations collected on mismatched hardware are worth less than the marketing around them suggests.

What none of these announcements delivered was the thing buyers actually need: a published task-success benchmark for sustained, unsupervised manipulation. Cycle life claims, repeatability figures and simulated transfer results are all useful, and all of them stop short of "this robot did this task a thousand times in a factory without help." That number is where the next round of competition will be decided, and it is the one number the IROS 2026 floor conspicuously did not have.

For more coverage of this space, see our robotics and drones section, and our recent reporting on humanoid deployment and embodied AI manufacturing.

Images

Illustration: a five-finger servo-electric robotic hand of the general class showcased at IROS 2026. Not a photograph of any specific product named in this article.

Illustration: an exposed prototype multi-finger robotic hand laid out beside a wired hand-shaped data glove and control electronics, the hardware-plus-teaching-interface pairing that recurred across IROS 2026 announcements.

IROS 2026 in Pittsburgh Made the Case That Robot Hands, Not Legs, Are the Bottleneck

Introduction

The Four-Finger Decision

Hardware and Its Teaching Interface

Fewer Fingers, More Sensors

The Volume Question

Conclusion

Images

References