Dexterous manipulation remains one of the biggest barriers keeping robots from successfully tackling a wide range of everyday tasks. A sense of touch could be the key, but a lack of quality data has held back progress. This is now starting to change as academic labs and startups race to build new tactile datasets and techniques to put them to use.

Over the last few years, vision-language-action (VLA) models have significantly improved the ability of robots to carry out complex tasks involving objects and environments they’ve never encountered before. Pretrained on huge amounts of images, video, and text, and then fine-tuned on a smaller number of teleoperated robot demonstrations, these VLAs can guide robots through a growing range of everyday jobs—like folding laundry, tidying living rooms, and even operating kitchen gadgets—using just a video feed and natural language instructions.

But robots still struggle with tasks that require fine-grained hand control, such as handling deformable materials or manipulating small objects—plugging in a USB cable or turning a key in a lock, for example. That’s partly because VLAs ignore one of the primary sources of information humans rely on in these situations: tactile feedback.

Manipulating Like Humans

“Most dexterous manipulation can be done by humans with their eyes closed,” says Trevor Darrell, professor of computer science at the University of California, Berkeley. “Understanding force, slip, and precise grasping is not something that can be done well with traditional vision sensors.”

However, making effective use of tactile sensors is difficult. Tactile sensor data has very different characteristics to the image data VLAs are normally trained on, and tactile datasets lag far behind the internet-scale of many vision and language datasets. To get around this, Darrell’s team devised a way to first pretrain a model on existing datasets before giving it a sense of touch by training a specialist submodel on 100 hours of specially collected, high-quality tactile data, including demonstrations of common actions like wiping, grasping, twisting, or pouring using more than 200 different household objects.

Explore an interactive visualizer of a small portion of the T-Rex dataset. T-Rex

Putting the tactile data to use was not straightforward. The goal was for a robot to be able to use the tactile signal to correct its grip in real time as it manipulated objects. But this requires reaction times faster than most vision-language models operate at. This mismatch is a significant challenge, Darrell says, so the team used separate submodels, known as “experts,” to handle high-level actions and low-level tactile control in a way that’s quick enough for the tactile feedback to be useful.

The action expert produces motion plans, while the tactile expert, which operates four times faster, uses tactile feedback to adjust the motion plan in real time based on what the robot is feeling as it goes. The model was then fine-tuned on about 100 teleoperated demonstrations of relatively complex manipulation tasks, such as screwing in a light bulb, applying toothpaste to a toothbrush, or transferring an egg between trays, where it averaged a success rate of 65 percent across 12 tasks—nearly double the best VLA model.

Data Diversity

One limitation, admits Darrell, is that his data comes from a single instance of robotic hardware. Robot hands range from fully articulated five-finger designs to simple pincer grippers, and tactile sensors can rely on fundamentally different physics, from measuring changes in resistance to recording images of a soft gel pad deforming. That makes most tactile AI research sensor-specific, says Chengbo Yuan, a master’s student at Tsinghua University in Beijing, and makes it hard to share data and transfer learnings between groups.

Yuan recently set out to tackle this problem by aggregating more than 3,000 hours of tactile robotic data from publicly available datasets, covering 21 sensor types and a variety of robot embodiments. Yuan says they were inspired by efforts like the Open X-Embodiment collaboration, which pooled data from many robots and led to models that generalize to hardware not used in training. Yuan’s team then designed a hardware-agnostic model that can train on this diverse data by converting each sensor’s output into a shared format and mapping it onto labeled positions on a template of a human hand. This model was much more successful than a baseline model, even on hardware it had never encountered before. Yuan puts that down to it acquiring “some kind of common sense of tactile knowledge,” by training on such diverse setups.

Chasing Scale

Despite the promising results, Yuan thinks more tactile data is needed, and his group is now leading an 80-institution collaboration to collate a larger set of teleoperated demonstrations using a standardized approach to tactile data collection and processing. In the meantime, Fudan University in Shanghai and its spin-out NeoteAI have already produced a tactile dataset an order of magnitude larger than previous efforts. Using a proprietary sensor attached to a variety of robotic arms and a handheld gripper operated by humans, they have collected more than 30,000 hours of demonstrations with synchronized visual and tactile data.

The researchers used this data to train a model that doesn’t just react to touch, but also proactively predicts what the robot should be feeling to help guide and assess actions, significantly improving performance. Shunlin Lu, a postdoc researcher at Fudan University and CTO of NeoteAI, says the results are clear evidence that access to large-scale and diverse tactile data leads to significant performance gains.

Robot manipulation policies with a tactile component offer improved performance on a variety of real-world tasks.NeoteAI

Another approach to scaling tactile data could be to piggyback on the vast quantities of visual robotics data already collected. Researchers at the University of Southern California, in Los Angeles, recently released a model that learned to infer tactile information from visual data, by training it on more than 2,700 demonstrations of everyday manipulation using a handheld gripper that records both tactile data and images from a camera on the device. The model learned associations between images of the gripper coming into contact with objects and the amount of pressure felt by the tactile sensors at that moment, giving even robots without tactile sensors a rudimentary sense of touch that the researchers showed to be particularly useful for contact-rich manipulation tasks. But their broader ambition is to use the generator to add tactile data to existing vision datasets.

How much tactile data will be required for breakthroughs in dexterous tasks remains unclear. So far, tactile training’s main contribution has been to make robots more efficient learners at tasks already within reach like picking and placing objects, says Yuan, and he suspects new algorithms may be required to tackle problems truly impossible without touch.

Long Cheng of the Chinese Academy of Sciences in Beijing also thinks raw data is no panacea. “Data is good,” he says. “But how to use them correctly is another issue.” The problem, he notes, is that vision provides a continuous, high-bandwidth stream of pixels, while tactile signals are sparse and intermittent, so models learn to ignore them. His solution, being presented at IROS 2026 later this month, is a model that predicts what a robot will feel from vision alone and then compares it against real tactile input. A large gap between the two means the sensor is detecting something the robot would otherwise miss, so these surprising signals are amplified while predictable ones are dampened. Across five contact-rich tasks, the approach averaged 62.8 percent success against 28.2 percent for the same model without touch.

Lu is more confident that data scaling could have similar benefits to those seen in areas like language and vision. He guesses closer to 100,000 hours, collected in varied, real-world settings rather than in the lab, could unlock new capabilities. Either way, the field now has some early signs that larger tactile datasets and smarter ways to use them can give robots a significant boost on some of the most challenging tasks. “I think tactile intelligence is actually the next step for physical AI,” Lu says.

X Square Robot is showcasing its full portfolio of embodied AI technologies at the World Robot Conference 2026 in Beijing, presenting applications ranging from logistics automation and dexterous manipulation to household robotics and AI training-data collection. The Shenzhen-based company is exhibiting at booth C107 during WRC, which runs from August 19 to 23, bringing together […]

Sony has filed a US patent application for a robotic hand featuring a human-like opposable thumb, providing another intriguing indication of the electronics giant’s continuing research into humanoid and embodied robotics. The application, US 2026/0225261 A1, was published by the US Patent and Trademark Office on August 6 and describes a robot hand designed to […]

Whatever happened to ASIMO? Honda’s famous humanoid robot (pictured below with what we believe is an actual human female) disappeared from public view several years ago, just before an extraordinary new wave of humanoid development swept through the robotics industry. But Honda didn’t abandon humanoid robotics. Instead, much of the technology and expertise accumulated through […]

A central challenge in robotics is building policies that generalize beyond the demonstrations they’re trained on. A policy that succeeds in a training scene...

A central challenge in robotics is building policies that generalize beyond the demonstrations they’re trained on. A policy that succeeds in a training scene often fails when object shapes, positions, or lighting change. Generalizing to these new conditions requires the policy to understand the tasks underlying physics, not just mimic the demonstrations. This ability comes from the backbone it’s…

Source

mimic robotics, a physical AI company developing Video-Action Models for industrial automation, has introduced FLUX-mimic – a next-generation Video-Action Model developed in collaboration with Black Forest Labs – that enables robots to learn and perform complex manipulation tasks in real-world industrial environments. FLUX-mimic combines mimic’s expertise in robot learning, dexterous manipulation, and production deployment with […]

This article is brought to you by AGILINK.

Throughout the exhibition hall at the 2026 IEEE International Conference on Robotics (ICRA), in Vienna, one demonstration seemed to attract a disproportionate amount of attention.

Two robotic hands were making a balloon dog. Slowly and deliberately, the robot twisted a long balloon into loops, bends, and joints without popping it. Visitors stopped, watched, and often returned with colleagues to watch again.

AGILINK’s balloon dog demonstration draws a crowd at ICRA 2026.AGILINK

At first glance, the demonstration appeared almost playful. Among roboticists, however, balloon twisting is widely recognized as an unusually difficult manipulation task.

A balloon is lightweight, highly deformable, slippery, and extremely sensitive to force. Every twist changes its geometry and internal pressure, turning a seemingly simple activity into a continuously changing physical interaction problem.

Humans navigate those changes almost intuitively. While making a balloon animal, people rarely think consciously about force regulation, slip prevention, or contact stability. They simply adjust.

For robots, those adjustments remain remarkably difficult. The challenge is not merely moving fingers to the right positions. The harder part is maintaining stable interaction while the object itself is changing.

Highlights from AGILINK’s ICRA 2026 demonstrations, including visuotactile sensing, in-hand manipulation, balloon-animal shaping, and other contact-rich tasks enabled by the company’s latest OmniHand platform.AGILINK

That distinction helps explain why the balloon dog drew so much attention in Vienna. What appeared to be a dexterity demonstration was, in many ways, a demonstration about contact itself.

As robotic manipulation continues to advance, a growing number of researchers are arriving at a similar conclusion: many of the hardest problems in robotics begin only after contact occurs.

Motion and Contact Intelligence for Robot Manipulation

Balloon twisting combines two challenges that robotics has traditionally struggled to solve simultaneously: long-horizon task execution and contact-rich manipulation.

The first concerns motion.

A balloon dog is not created through a single grasp or twist. It emerges through a carefully ordered sequence of manipulations, each setting the conditions for what follows. A small rotational error introduced early may appear insignificant at first, yet several steps later it can prevent the final structure from forming altogether.

In that sense, balloon twisting is a long-horizon task. Success depends not only on performing individual actions correctly, but also on preserving the future feasibility of the entire manipulation process.

To address this challenge, AGILINK began by collecting demonstrations from professional balloon artists. Human actions were mapped onto robotic hands to establish an initial manipulation policy. But successful demonstrations alone were insufficient.

In practice, some of the most valuable learning occurred when execution began to drift toward failure. Whenever instability emerged, human operators intervened and corrected the manipulation in real time. Those interventions were recorded and incorporated into reinforcement-learning cycles, allowing the system to learn not only how successful demonstrations unfold, but also how experienced operators recover when things start to go wrong.

Through this process, the robot gradually acquired the capabilities required for long-horizon task execution—a collection of abilities that AGILINK groups under the term motion intelligence: the ability to generate actions, coordinate bimanual behaviors, and execute extended manipulation sequences under real-world uncertainty.

OmniHand 3 Ultra-M on display at ICRA 2026.AGILINK

Yet motion alone does not explain why balloon twisting remains difficult. The second challenge is contact.

The robot must continuously regulate force, adjust contact locations, and respond to subtle changes in the object’s state. These decisions are difficult to encode through explicit rules. Even skilled human operators often rely on tactile intuition developed through experience rather than consciously articulated strategies.

Analysis of those interventions revealed that many failures did not originate from incorrect action sequences, but from the breakdown of contact itself.

To better capture those interaction dynamics, AGILINK collected contact-centric intervention data and incorporated those interactions into reinforcement-learning training. Rather than learning only which motions to perform, the system also learned how humans maintain stability when contact conditions begin to deteriorate.

AGILINK describes this capability as contact intelligence: the ability to establish, maintain, and adapt physical interaction as force distribution, friction, deformation, and contact geometry continuously evolve.

The distinction between the two capabilities is subtle but important. Motion intelligence determines what the robot intends to do. Contact intelligence determines whether it can continue doing it. For balloon twisting, both are necessary. One provides the sequence of actions. The other keeps those actions physically viable.

YouTuber KhanFlicks follows OmniHand’s motions while learning to fold a balloon dog at the AGILINK booth.AGILINK

Between a balloon slipping away and a balloon bursting lies a narrow region of stability. Successful manipulation depends on finding that region—and remaining within it throughout the task.

Introducing the OmniHand 3 Ultra-M Dexterous Hand

The balloon dog demonstration showcased a manipulation capability. It also revealed a broader question. How much contact intelligence can be achieved through learning alone? A robot can only regulate what it can perceive. It can only respond as quickly as its hardware allows.

As manipulation tasks become increasingly complex, researchers are finding that progress depends not only on better policies, but also on richer sensing and faster physical response.

That realization formed the backdrop for AGILINK’s second major announcement at ICRA 2026. Alongside the balloon dog demonstration, the company introduced the OmniHand 3 Ultra-M.

OmniHand 3 Ultra-M closely matches the size of an adult human hand.AGILINK

The two exhibits represented different stages of the same technological trajectory. If the balloon dog demonstrated what contact intelligence can already accomplish today, Ultra-M was designed to explore what contact intelligence may require next.

Building Hardware for Contact Intelligence

Roughly the size of an adult human hand, the OmniHand 3 Ultra-M integrates 20 active degrees of freedom within a human-scale form factor.

Its most distinctive feature is a fully direct-drive architecture. By adopting direct-drive actuation throughout the system, the hand is designed to enable faster and more transparent force regulation and higher force-control bandwidth, enabling faster response as contact conditions change. For contact-rich manipulation, responsiveness can be as important as sensing itself.

By adopting direct-drive actuation throughout the system, the OmniHand 3 Ultra-M is designed to enable faster and more transparent force regulation and higher force-control bandwidth, enabling faster response as contact conditions change.

The platform also incorporates tactile sensing across nearly the entire hand. Each fingertip contains a miniature vision-based tactile sensor, while more than 300 three-dimensional tactile sensing points are distributed throughout the palm. Together, they provide information not only about where contact occurs, but how contact is evolving.

The system is designed to estimate pressure distribution, shear forces, local deformation, slip tendencies, and other interaction dynamics that often remain invisible to conventional position-based control systems.

According to AGILINK’s tests, individual sensors achieve force resolution of approximately 0.005 N—roughly equivalent to detecting the weight of a sheet of paper resting on a fingertip. Spatial resolution reaches approximately 0.04 mm, while sensing density approaches 50,000 sensing points per square centimeter.

OmniHand 3 Ultra-M recognizes feather texture through vision-based tactile sensing.AGILINK

For dexterous robots, contact has traditionally been a largely hidden process. Ultra-M is designed to make that process more observable.

Rather than simply detecting that contact has occurred, the system attempts to resolve where interaction is happening, how forces are distributed, whether instability is beginning to emerge, and how manipulation strategies should adapt in response.

The balloon dog offered a glimpse of what contact intelligence can already accomplish. Ultra-M explores a different question: what capabilities may be required to push contact intelligence further?

The Physical World Remains the Hardest Benchmark

The significance of contact intelligence extends far beyond balloon animals. Many tasks that continue to resist automation involve unstable or deformable interaction: cable insertion, garment handling, flexible packaging, delicate assembly, connector mating, tool use, and household manipulation.

These tasks are difficult not because robots cannot reach the correct location, but because maintaining stable interaction after contact begins remains extraordinarily hard.

For decades, robotics achieved many of its successes by reducing uncertainty. Factories were engineered to make robotic motion predictable, repeatable, and highly structured. The physical world behaves differently.

A growing share of robotics research is shifting toward interaction itself—understanding how robots can establish, maintain, and adapt physical contact within environments that remain fundamentally unpredictable.

Objects shift. Materials deform. Friction changes. Contact evolves. Real environments rarely follow scripts. Seen through that lens, the balloon dog was never really about the balloon dog. What attracted attention at ICRA was not simply a visually impressive demonstration, but what it revealed: intelligence in the physical world is ultimately measured through interaction.

As motion generation continues to mature, a growing share of robotics research is shifting toward interaction itself—understanding how robots can establish, maintain, and adapt physical contact within environments that remain fundamentally unpredictable.

For robots moving beyond structured environments and into less predictable real-world settings, managing contact may become as important as motion itself.

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

ICRA 2026: 1–5 June 2026, VIENNA

RSS 2026: 13–17 July 2026, SYDNEY

Summer School on Multi-Robot Systems: 29 July–4 August 2026, PRAGUE

Actuate 2026: 18–19 August 2026, SAN FRANCISCO

Enjoy today’s videos!

Introducing GENE-26.5—the first AI brain to give robots human-level physical manipulation capabilities. Cooking a full meal. Cracking an egg one-handed. Conducting lab experiments. Wire harnessing. Even playing the piano. Tasks that were impossible for robots. Until now.

[ Genesis AI ] via [ TechCrunch ]

This is Labububot—one of the rarest monsters on Earth. Twelve Labubu heads are reconstituted into a single spherical form: a Frankenstein’s Monster of pop culture iconography. Labububot is a playful critique of social robots, and a question made physical—what do the monsters we make reveal about the monsters we are?

[ MIT Media Lab ]

Watch Spot crouch, jump, climb boxes, and leap across gaps, controlled by a neural network trained with reinforcement learning (RL) and multi-expert distillation.

[ Robotics and AI Institute ]

Good, now there is a robot that can take over exercise for me.

[ Kepler ]

Additive manufacturing has become an enabling technology, but existing techniques are not capable of directly 3D printing high-current electromagnetic actuators due to material and design limitations. In this work, a novel 3D-printable, multilayer, wave-winding topology is created for high-efficiency electric motors.

[ Sensing Technologies Laboratory ]

NASA is pushing the limits of flight on Mars—by spinning helicopter rotor blades so fast, they’re breaking the sound barrier. During recent tests at NASA’s Jet Propulsion Laboratory, engineers accelerated the tips of next-generation rotor blades beyond Mach 1 inside a special chamber that simulates the atmospheric conditions of the Red Planet.

[ NASA Jet Propulsion Laboratory ]

Balancing commercial goals and robotics research can be tricky, but with Atlas, we’re making it work.

[ Boston Dynamics ]

Open Duck Mini is an open-source version of Disney’s BDX droids, and you can play with it in your browser.

[ Open Duck Mini Viewer ]

Thanks, Masato!

Automated inspection of steel structures using magnetic climbing robots can reduce costs and improve safety, but many such structures feature interior corners that are challenging for wheeled or tracked robots to traverse. We present the first magnetic-wheeled robot to use X-ray fluorescence for steel structure inspection, Sally, capable of overcoming all interior corner transition types, traversing small obstacles, and maneuvering in tight spaces.

[ Robomechanics Lab ]

I don’t know what this is, but it’s coming soon from SwitchBot.

[ SwitchBot ]

You probably know the answers to these questions already, but this ELI5 from Aaron Ames is still fun.

[ Wired ]

Jim Fan, who leads the embodied autonomous research group at Nvidia, returns to AI Ascent to argue that robotics is entering its endgame—and that the playbook is already written.

[ Sequoia ]

Video Friday is your weekly selection of awesome robotics videos, collected by your friends at IEEE Spectrum robotics. We also post a weekly calendar of upcoming robotics events for the next few months. Please send us your events for inclusion.

ICRA 2026: 1–5 June 2026, VIENNA

RSS 2026: 13–17 July 2026, SYDNEY

Summer School on Multi-Robot Systems: 29 July–4 August 2026, PRAGUE

Actuate 2026: 18–19 August 2026, SAN FRANCISCO

Enjoy today’s videos!

Figure is now able to produce 55 robots per week, which will be “allocated to internal research and development groups, data collection, efforts for robots to perform end-to-end housework, and commercial use-case development.” Er, that seems like a lot of robots to be making when commercial use cases are still “in development,” doesn’t it?

[ Figure ]

The opening of the NEO Factory in Hayward, Calif., marks a fundamental shift in humanoid robotics: The United States’ most vertically integrated robot factory has now begun full-scale production, bringing end-to-end manufacturing of NEO under one roof. Spanning 58,000 square feet and employing over 200 team members, 1X designs and builds every critical component in-house—motors, batteries, transmissions, sensors, structures, and final assembly—enabling faster iteration, superior safety, and true American scale. With the first robots already coming off the line and consumer shipments planned for 2026, this is the critical milestone that turns the vision of abundant, general-purpose home robots into reality.

Scale will fix everything...?

[ 1X ]

Unlike statically stable robots, a dynamically balanced robot can shift its center of mass to accommodate loads without tipping over, so we like to see just how far we can push our software. Getting Digit to stand on one leg pushes the limits of our sim-to-real pipeline training methodologies—even the slightest model mismatches can lead to instability.

[ Agility ]

In this work, we develop a tactile-enabled whole-body humanoid manipulation system for stable, dexterous, contact-rich real-world manipulation. Our system combines VR-based whole-body teleoperation, a lower-body controller based on reinforced learning, dexterous hand retargeting, distributed tactile sensing, and a multimodal policy called Humanoid Transformer with Touch Dreaming (HTD).

[ Humanoid Touch Dream ]

Thanks, Yaru!

Originally posted two years ago, “Can I Have a Pet T. Rex?” is a short interdisciplinary portrait documentary. It features paleontologist and Kod*lab postdoc Aja Mia Carter and the Kod*lab robotics researchers Wei-Hsi Chen (also a postdoc) and J. Diego Caporale, a Ph.D. student.

It’s been two years! Where is her pet T. rex!?

[ Kod*Lab ]

I am not entirely sure why CMU and HEBI had robots at the 2026 NFL Draft, but I’m entirely sure that it made it more interesting to watch.

[ HEBI Robotics ]

Thanks, Trevor!

Ethan Lauer, a software engineer, answers your questions about robot perception, world modeling, and what spooks our Stretch robot.

[ Boston Dynamics ]

Yet another thing that a robot is consistently better at than I am.

[ Generalist ]

If you’re wondering where all those reported humanoid robot sales are coming from, it’s because every big company needs one or two for this sort of thing.

[ Impress ]

Full-color laser yo-yo zapper, a phrase never before written in the history of the universe.

[ Ishikawa Group Laboratory ]

The future of the L’Oréal Pro 2026 Le Hair Show is...a bald robot?

[ LimX Dynamics ]

Meet MagicHand H01, our all-new dexterous hand.

[ MagicLab ]

This is briefly one of the flattest quadrupeds I have ever seen.

[ DEEP Robotics ]

I appreciate that Engineered Arts did not try to cover up the sound in this video.

[ Engineered Arts ]

This is very impressive considering that magnets are basically indistinguishable from magic.

[ Sung Lab ]

NASA has two rovers on Mars, but they’re exploring entirely different eras of the planet’s past. Separated by 2,300 miles, the two rovers are uncovering clues from very different moments in Martian history. Perseverance is on the rim of Jezero Crater, where it’s studying some of the oldest Martian terrain ever explored while searching for signs of ancient microbial life. Meanwhile, Curiosity is climbing Mount Sharp inside Gale Crater, where layers of rock reveal how Mars’s climate changed as water dried up from its surface.

[ NASA ]

We’ve built a surgical robot to automate key steps in the process of receiving a Neuralink implant to promote safety, reliability, and scalability.

[ Neuralink ]

The Chinese-made Unitree G1 humanoid robots are making their way into the United States. And they aren’t just in viral videos but in major tech companies like OpenAI and Nvidia, and top academic institutions. Most arrive through Robostore, a robotics reseller based on Long Island. I went there to watch them come off the pallet, then brought one to my home to see what it could actually do. Are these the future of home robots? A security risk? A Chinese surveillance system on legs? I got answers—and a broken toe.

[ New Things ]

How do autonomous robots make decisions when the world is unpredictable? From self-driving cars to drone swarms, autonomous systems must operate under uncertainty—making real-time decisions with incomplete or unreliable data. In this video, Harvard SEAS Prof. Stephanie Gil explains how AI-powered robots coordinate, adapt, and stay safe in complex, real-world environments.

[ Harvard University ]