This text is dropped at you by AGILINK.
All through the exhibition corridor on the 2026 IEEE Worldwide Convention on Robotics (ICRA), in Vienna, one demonstration appeared to draw a disproportionate quantity of consideration.
Two robotic palms had been making a balloon canine. Slowly and intentionally, the robotic twisted a protracted balloon into loops, bends, and joints with out popping it. Guests stopped, watched, and infrequently returned with colleagues to look at once more.
AGILINK’s balloon canine demonstration attracts a crowd at ICRA 2026.AGILINK
At first look, the demonstration appeared nearly playful. Amongst roboticists, nonetheless, balloon twisting is widely known as an unusually tough manipulation job.
A balloon is light-weight, extremely deformable, slippery, and intensely delicate to power. Each twist modifications its geometry and inner strain, turning a seemingly easy exercise right into a constantly altering bodily interplay drawback.
People navigate these modifications nearly intuitively. Whereas making a balloon animal, individuals hardly ever suppose consciously about power regulation, slip prevention, or contact stability. They merely regulate.
For robots, these changes stay remarkably tough. The problem is just not merely transferring fingers to the precise positions. The tougher half is sustaining steady interplay whereas the item itself is altering.
Highlights from AGILINK’s ICRA 2026 demonstrations, together with visuotactile sensing, in-hand manipulation, balloon-animal shaping, and different contact-rich duties enabled by the corporate’s newest OmniHand platform.AGILINK
That distinction helps clarify why the balloon canine drew a lot consideration in Vienna. What seemed to be a dexterity demonstration was, in some ways, an illustration about contact itself.
As robotic manipulation continues to advance, a rising variety of researchers are arriving at an identical conclusion: lots of the hardest issues in robotics start solely after contact happens.
Balloon twisting combines two challenges that robotics has historically struggled to resolve concurrently: long-horizon job execution and contact-rich manipulation.
The primary issues movement.
A balloon canine is just not created via a single grasp or twist. It emerges via a rigorously ordered sequence of manipulations, every setting the situations for what follows. A small rotational error launched early might seem insignificant at first, but a number of steps later it will possibly forestall the ultimate construction from forming altogether.
In that sense, balloon twisting is a long-horizon job. Success relies upon not solely on performing particular person actions accurately, but additionally on preserving the longer term feasibility of the whole manipulation course of.
To deal with this problem, AGILINK started by accumulating demonstrations from skilled balloon artists. Human actions had been mapped onto robotic palms to determine an preliminary manipulation coverage. However profitable demonstrations alone had been inadequate.
In follow, a few of the Most worthy studying occurred when execution started to float towards failure. Each time instability emerged, human operators intervened and corrected the manipulation in actual time. These interventions had been recorded and integrated into reinforcement-learning cycles, permitting the system to be taught not solely how profitable demonstrations unfold, but additionally how skilled operators get well when issues begin to go flawed.
By way of this course of, the robotic steadily acquired the capabilities required for long-horizon job execution—a group of talents that AGILINK teams beneath the time period movement intelligence: the power to generate actions, coordinate bimanual behaviors, and execute prolonged manipulation sequences beneath real-world uncertainty.
OmniHand 3 Extremely-M on show at ICRA 2026.AGILINK
But movement alone doesn’t clarify why balloon twisting stays tough. The second problem is contact.
The robotic should constantly regulate power, regulate contact areas, and reply to refined modifications within the object’s state. These selections are tough to encode via express guidelines. Even expert human operators typically depend on tactile instinct developed via expertise relatively than consciously articulated methods.
Evaluation of these interventions revealed that many failures didn’t originate from incorrect motion sequences, however from the breakdown of contact itself.
To higher seize these interplay dynamics, AGILINK collected contact-centric intervention information and integrated these interactions into reinforcement-learning coaching. Reasonably than studying solely which motions to carry out, the system additionally discovered how people preserve stability when contact situations start to deteriorate.
AGILINK describes this functionality as contact intelligence: the power to determine, preserve, and adapt bodily interplay as power distribution, friction, deformation, and make contact with geometry constantly evolve.
The excellence between the 2 capabilities is refined however essential. Movement intelligence determines what the robotic intends to do. Contact intelligence determines whether or not it will possibly proceed doing it. For balloon twisting, each are vital. One supplies the sequence of actions. The opposite retains these actions bodily viable.
YouTuber KhanFlicks follows OmniHand’s motions whereas studying to fold a balloon canine on the AGILINK sales space.AGILINK
Between a balloon slipping away and a balloon bursting lies a slender area of stability. Profitable manipulation is dependent upon discovering that area—and remaining inside it all through the duty.
Introducing the OmniHand 3 Extremely-M Dexterous Hand
The balloon canine demonstration showcased a manipulation functionality. It additionally revealed a broader query. How a lot contact intelligence might be achieved via studying alone? A robotic can solely regulate what it will possibly understand. It might solely reply as shortly as its {hardware} permits.
As manipulation duties develop into more and more complicated, researchers are discovering that progress relies upon not solely on higher insurance policies, but additionally on richer sensing and quicker bodily response.
That realization fashioned the backdrop for AGILINK’s second main announcement at ICRA 2026. Alongside the balloon canine demonstration, the corporate launched the OmniHand 3 Ultra-M.
OmniHand 3 Extremely-M intently matches the scale of an grownup human hand.AGILINK
The 2 reveals represented completely different levels of the identical technological trajectory. If the balloon canine demonstrated what contact intelligence can already accomplish immediately, Extremely-M was designed to discover what contact intelligence might require subsequent.
Constructing {Hardware} for Contact Intelligence
Roughly the scale of an grownup human hand, the OmniHand 3 Ultra-M integrates 20 active degrees of freedom inside a human-scale kind issue.
Its most distinctive function is a completely direct-drive structure. By adopting direct-drive actuation all through the system, the hand is designed to allow quicker and extra clear power regulation and better force-control bandwidth, enabling quicker response as contact situations change. For contact-rich manipulation, responsiveness might be as essential as sensing itself.
By adopting direct-drive actuation all through the system, the OmniHand 3 Extremely-M is designed to allow quicker and extra clear power regulation and better force-control bandwidth, enabling quicker response as contact situations change.
The platform additionally incorporates tactile sensing throughout almost the whole hand. Every fingertip incorporates a miniature vision-based tactile sensor, whereas greater than 300 three-dimensional tactile sensing factors are distributed all through the palm. Collectively, they supply info not solely about the place contact happens, however how contact is evolving.
The system is designed to estimate strain distribution, shear forces, native deformation, slip tendencies, and different interplay dynamics that always stay invisible to standard position-based management programs.
In line with AGILINK’s assessments, particular person sensors obtain power decision of roughly 0.005 N—roughly equal to detecting the burden of a sheet of paper resting on a fingertip. Spatial decision reaches roughly 0.04 mm, whereas sensing density approaches 50,000 sensing factors per sq. centimeter.
OmniHand 3 Extremely-M acknowledges feather texture via vision-based tactile sensing.AGILINK
For dexterous robots, contact has historically been a largely hidden course of. Extremely-M is designed to make that course of extra observable.
Reasonably than merely detecting that contact has occurred, the system makes an attempt to resolve the place interplay is going on, how forces are distributed, whether or not instability is starting to emerge, and the way manipulation methods ought to adapt in response.
The balloon canine supplied a glimpse of what contact intelligence can already accomplish. Extremely-M explores a unique query: what capabilities could also be required to push contact intelligence additional?
The Bodily World Stays the Hardest Benchmark
The importance of contact intelligence extends far past balloon animals. Many duties that proceed to withstand automation contain unstable or deformable interplay: cable insertion, garment dealing with, versatile packaging, delicate meeting, connector mating, instrument use, and family manipulation.
These duties are tough not as a result of robots can’t attain the proper location, however as a result of sustaining steady interplay after contact begins stays terribly laborious.
For many years, robotics achieved a lot of its successes by lowering uncertainty. Factories had been engineered to make robotic movement predictable, repeatable, and extremely structured. The bodily world behaves otherwise.
A rising share of robotics analysis is shifting towards interplay itself—understanding how robots can set up, preserve, and adapt bodily contact inside environments that stay essentially unpredictable.
Objects shift. Supplies deform. Friction modifications. Contact evolves. Actual environments hardly ever comply with scripts. Seen via that lens, the balloon canine was by no means actually in regards to the balloon canine. What attracted consideration at ICRA was not merely a visually spectacular demonstration, however what it revealed: intelligence within the bodily world is finally measured via interplay.
As movement technology continues to mature, a rising share of robotics analysis is shifting towards interplay itself—understanding how robots can set up, preserve, and adapt bodily contact inside environments that stay essentially unpredictable.
For robots transferring past structured environments and into much less predictable real-world settings, managing contact might develop into as essential as movement itself.
