Despite the recent deluge of videos of humanoid robots doing backflips and kung-fu, the actual pace of progress towards humanoids that can do economically viable jobs at scale has at times seemed a bit plodding. There are some good reasons for this: a humanoid robot needs to be powerful enough to do useful work, safe enough for humans to walk past, and have enough battery life to limit how much of the day it’s hooked to a charger. But all three of those necessary features actively work against each other.

With its new Digit 5 robot, announced today, Agility Robotics may have found the sweet spot. Digit 5 is a humanoid worker that can lift 23 kilograms up 2.1 meters, can work in close proximity to people without relying on physical barriers, and can operate for more than 20 hours a day. This combination of power, safety, and battery life has resulted in a 1.8 m tall, 129 kg robot that prioritizes functionality over style and pretty much everything else, but this is exactly what it’s going to take to get humanoids a job.

Humanoid robot companies have been reluctant to talk about safety, because there are no easy answers to “what happens if your fundamentally unstable bipedal robot falls over onto me/my pet/my baby?” And there are still no easy answers, but Agility does have an answer that allows Digit to be verifiably safe. That answer is to make sure that it’s physically impossible for Digit to fall over on anyone, ever, by “autonomously avoiding, stopping or assuming a seated position,” according to today’s press release. In other words, as a person approaches Digit, Digit will put down whatever it’s carrying and then if necessary make sure that it’s stably seated on the ground before that person gets near. It may not be elegant, but it works, which is more than can be said for any other commercial humanoid that I’m aware of.

The other big change is Digit’s legs. Somewhat famously, Agility robotics introduced first Cassie and then a whole series of Digit robots featuring bird-like ‘backwards’ legs. In his 2019 article for IEEE Spectrum, Agility co-founder Jonathan Hurst explained that this leg configuration was the result of a careful analysis of the physics of animal locomotion, rather than for the leg to look like any sort of animal in particular. The unfortunate reality for those of us who were fans of Cassie is that bird legs are optimal for dynamic motion, while human legs are better for squats and lifts, which will make sense to anyone who has seen an ostrich trying to lift a heavy box up off the floor and also anyone who has tried to outrun a cassowary.

How much will Digit 5 cost?

We also know a couple of things about Digit 5 that weren’t included in today’s announcement, thanks to Agility’s June SEC filing in advance of their plan to go public by the end of 2026. For example, at launch, Digit 5’s BOM (bill of materials) cost is likely to be somewhere between $150,000 and $200,000. This is just the cost of the parts that make up a Digit robot, not what it costs for Agility to actually build one. Based on real production data, Agility is anticipating that with some near-term optimizing and at a volume of 10,000 units per year, that cost should drop to under $50,000 per robot.

In this rendering, Digit 5 robots work in a factory with humans nearby, no safety barriers needed. When a human approaches, the robot puts down what it’s carrying and squats down so it can’t fall on anybody.Agility

Arguably less important than the per-robot cost is how Agility (and its customers) will profit from Digit 5. In the filing, Agility estimates (using “rounded estimates” which are “purely illustrative”) that Digit 5 robots will be offered to customers as a service at something like $8,500 per month. Based on 20 hours per day of work, and assuming that the total cost of a human worker to their employer is $30.50/hr, Digit 5 as a service could save employers $100,000 per year, per robot.

It’s important to note that all of these numbers are just estimates, and that all kinds of things (many of them not under Agility’s control) could cause them to change. They’re most useful as an illustration of Agility’s broad approach to making Digit 5 profitable. It’s also important to note the assumption here with such a direct comparison is that Digit 5 will be a more or less effortless drop-in replacement for human labor, which is not something that I’m entirely sure has ever happened with any robot, anywhere.

Where does Digit 5 go from here? Today’s press release says that “as of May 2026, Agility had more than $300 million in multi-year customer orders for Digit 5, with a sales pipeline of prospective customers across manufacturing, warehousing, and logistics.” That works out to comfortably under 1,000 robots, one-tenth of the full capacity of Agility’s RoboFab factory in Oregon.

Between now and all those robots, however, lies an unpredictable entry into the stock market through what’s called a SPAC merger. That’s expected to close within the next few months. Agility hopes to raise more than $620 million through the merger, and it will primarily spend the money to scale production of Digit 5, and get it to customers. Those in the EU and the UK should be able to buy the bots in 2027.

Agility Robotics has unveiled Digit 5, the next version of its humanoid robot for warehouses and factories.

The article Agility Robotics says its new Digit 5 robot can work next to people without safety fences appeared first on The Decoder.

The round for the two-year-old startup is coming together months after Mecka announced its Series A.

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.

Maven Robotics emerged from stealth today with a $100 million Series A and active deployments.

You sit down and put your arm in the cradle. You press a button. The machine takes it from there.

A near-infrared light sweeps your inner elbow, hunting for a vein. A puff of alcohol hits your skin. An ultrasound probe glides across your arm, mapping how deep the vessel runs and which way it bends. Doppler captures the direction of blood flow to rule out the artery.

The cuff tightens around your upper arm. The needle comes down and pierces the skin. Your blood flows into the collection tubes, each one tipped end over end nine times—no more, no less. The needle withdraws. You get a bandage. No human ever touched you.

This is what it’s like to have blood taken by Aletta, the first autonomous blood-draw device authorized for use in the United States. Developed by the Dutch medical robotics firm Vitestro, the system combines imaging technologies with advanced robotics and AI to do by algorithm what a human phlebotomist—a trained health care professional who finds veins and draws blood by hand—does by feel.

The U.S. Food and Drug Administration gave Aletta the go-ahead on 19 August for use on adults in nonhospitalized settings. The decision follows the lead of European regulators, who authorized the device two years earlier.

“You have to tip your cap to them,” says Max Balter, a surgical-robotics specialist at Medtronic who worked on autonomous blood-draw systems during grad school in the mid-2010s. “The engineering that they have is incredible…and with their FDA clearance, it moves the whole industry forward.”

Aletta Boosts Lab Capacity Amid Shortages

In a clinical trial involving more than 1,600 people in the Netherlands, Aletta successfully drew blood on the first attempt in 94.5 percent of cases, even among those with hard-to-access veins, people with obesity, and the elderly. When Aletta failed to identify a suitable vein, the patient was referred for conventional phlebotomy.

“It’s exceptional performance,” says Joe El-Khoury, a clinical chemist at Yale who was not involved in the Dutch trial. “It’s definitely as good if not better” than a typical professional phlebotomist.

Complications were minimal, with multiple built-in safeguards to detect problems, such as sensors that track arm movement and needle position, and to halt the process should something go awry. And for those whose veins proved too challenging, a phlebotomist remains on hand to take over when needed.

Notably, because a single phlebotomist can supervise up to three Aletta machines, the system should go a long way toward “helping clinical labs address the critical operational challenges related to the staffing shortages of phlebotomists,” says Luuk Giesen, chief medical officer of Vitestro.

That’s no small challenge in a profession with a median annual turnover rate of nearly 25 percent and a vacancy rate of close to 10 percent, according to surveys of medical laboratories that draw mostly from U.S. institutions. The resulting staffing shortages can limit labs’ capacity to meet demand for routine diagnostic testing of blood counts, cholesterol, metabolic markers, and more. Aletta could address that bottleneck.

Addressing Skin Tone Bias in Blood Draw AI

The promise of greater capacity, however, comes with a caveat: Aletta still fails in roughly one case out of 20. Who are those people?

Some may simply have elusive or unusually deep veins. But a more significant obstacle could be skin pigmentation. In particular, the melanin in darker skin can interfere with the near-infrared light Aletta uses to first map the veins near the skin’s surface and identify promising puncture sites. The technique relies on hemoglobin absorbing the light differently from surrounding tissue, and darker skin tones can absorb more of that light before it reaches the camera, thereby reducing the contrast.

Some experts have raised concerns that Aletta’s infrared sensors will perform poorly for people with darker skin tones, but Vitestro says that its machine also includes an ultrasound sensor, in part to mitigate that risk.Vitestro

Giesen recognizes that the issue could affect first-pass imaging, but notes that the main determinant of vein selection and needle placement is the ultrasound system, which relies on sound rather than light and should not be affected by skin pigmentation in the same way. “Ultrasound is skin-tone agnostic,” he says, adding that Vitestro has unpublished data showing no effect of skin tone on the system’s performance.

The company thus claims on its website that the “technology works well for all skin tones,” an assertion echoed in the FDA press release announcing the authorization of Vitestro’s device.

But given the history of racial disparities in medical devices—particularly optical technologies such as pulse oximeters, which can be less accurate in people with darker skin and went largely unrecognized as a problem for decades—such claims warrant evidence, says El-Khoury, who has written about the issue.

Brooke Katzman, a clinical chemist at the Mayo Clinic who is collaborating with Vitestro, also wants more evidence that samples collected by the robot are as suitable for testing as those drawn by hand.

The Dutch trial reported little damage to red blood cells, but other measures of sample quality, including clotted tubes, insufficient blood, and proper tube filling, still need to be assessed, as do the results of routine laboratory tests themselves. Katzman plans to launch a U.S.-based trial next year to collect just that sort of data.

“We’re going to do our due diligence,” she says. “Like any instrument we would bring into the lab, we’re going to put it through its paces before using it clinically.”

The Future of Automated Blood Testing

Aletta takes its name from the 19th-century physician Aletta Jacobs, the first female doctor in the Netherlands and founder of what is widely considered the world’s first birth-control clinic.

That nod to history is fitting for a technology that builds on decades of research in robotic phlebotomy by groups in Europe, the United States, and China, and MagicNurse. Yet few pushed the concept as far as biomedical engineer Martin Yarmush of Rutgers University in New Jersey, in whose lab Medtronic’s Balter completed his Ph.D.

In one version of their platform, the Rutgers team even coupled their robot to a benchtop blood analyzer, allowing it to draw samples and then measure levels of infection-fighting immune cells and oxygen-carrying red blood cells—all within minutes.

That all-in system never made it out of laboratory testing. And VascuLogic, the company spun out to commercialize the platform, is long defunct—though others, including ROPHAI, BHealthCare, and MagicNurse, continue to work in the space. But the Rutgers proof-of-concept demonstration points toward the tantalizing possibility of fully automated blood testing at the point of care, with robots handling everything from the needle stick to the analysis.

It is, in some ways, the promise that Theranos made—but built on conventional, validated laboratory technology rather than the dubious science and deception that brought that particular company down.

“I have no doubt that is the future,” says Gregory Retzinger, a clinical pathologist at the Northwestern University Feinberg School of Medicine, in Chicago, who collaborates with Vitestro and has tried the Aletta device himself. (“It was painless, it was fast,” he says.)

For now, Giesen says Vitestro is keeping its ambitions—and its machine—focused on the blood-collection process itself, though he believes Aletta could ultimately do far more. The company plans to launch Aletta in Europe next year, with the U.S. market to follow.

Imagine you are trapped under rubble after an earthquake and you see an electronics-covered cockroach with a spring-loaded needle on its back scuttling toward you. Although the sight might be unnerving, to say the least, this prototype paramedic cyborg, or “Paraborg,” might one day help deliver lifesaving aid to disaster victims who might be otherwise impossible to reach.

---

The Hardest Problems in Robotics

For decades, scientists have sought to develop cyborg insects as “a shortcut around some of the hardest problems in robotics,” says T. Thang Vo-Doan, director of the University of Queensland’s Biorobotics Lab in Brisbane, Australia, which just published a paper on the Paraborgs.

The University of Queensland

Building an insect-size robot “that can move reliably through rubble, climb over irregular surfaces, recover from falls, carry its own power, and still have room for useful sensors is extraordinarily difficult,” Vo-Doan says. An insect already comes with much of that mobility built in, so instead of trying to create artificial versions of every part of an insect’s body from scratch, researchers can graft an electronic interface onto an insect to make use of its existing capabilities.

Previously, scientists have shown they could steer cyborg insects such as beetles and moths. This prior work largely focused on controlling their movements to serve as passive sensor platforms.

In 2023, as Vo-Doan and fellow researcher Thanh Nho Do were talking about search-and-rescue cyborg insects shortly before that year’s IEEE International Conference on Robotics and Automation (ICRA), they asked, “What happens after an insect finds a trapped victim?” Vo-Doan recalls. “Could it go beyond locating someone and actually provide some form of assistance while rescuers are still trying to reach them?”

Giant Cockroaches to the Rescue

To answer this question, the roboticists experimented with giant burrowing cockroaches (Macropanesthia rhinoceros), which are native to Australia. The researchers needed an insect capable of carrying a large payload (over half its weight), and at roughly 40 grams in size, this species is the world’s heaviest species of roach. A larger insect is also easier to operate on to implant cybernetic interfaces.

The cockroaches were saddled with lightweight electronics that included electrodes implanted into both their antennae and small tail-like appendages known as cerci. Wirelessly activating these electrodes with a handheld gaming controller could steer the roaches left or right, spur them forward, or stop them from moving.

The insects were also equipped with either a wireless camera or a remote-controlled injector, which used a spring to launch a drug-filled syringe at a nearby target. A chemical reaction inside the syringe then generated a puff of carbon dioxide, which exerted pressure within the syringe to inject its payload into a target.

Paraborgs are designed to work in teams, with some carrying cameras and others carrying injectors with potentially lifesaving medications.The University of Queensland

The scientists decided not to load both a camera and an injector onto a single roach because the combined weight and bulk could impair their mobility on complex terrain. Having both sets of electronics would also increase energy demands, resulting in reduced operation time. Instead, the researchers envision a swarm approach with the Paraborgs, with different specialized cyborgs performing complementary roles.

In proof-of-concept tests, the scientists were able to successfully navigate the Paraborgs over a 2.5-meter course past three checkpoints before launching their needles at an 8-by-10-centimeter silicone target. In 25 trials, the cockroaches completed the course every single time and succeeded at injecting the target 72 percent of the time. “The long-term goal is to combine the insect’s advanced locomotion with sensing and intervention capabilities so we can reach and help more people, more quickly,” Vo-Doan says.

While the Paraborgs can be steered remotely, the cockroaches themselves are still very much alive and able to use their skills as bugs to navigate through complex terrain.The University of Queensland

The researchers acknowledge that “for someone who is already trapped or injured, seeing a cyborg insect approaching could understandably be a little surprising or unsettling at first,” Vo-Doan says. Ways to make it clear these insects were part of rescue efforts might include flashing lights, recognizable emergency markings, “or perhaps a tiny speaker delivering a simple message such as, ‘Help is on the way,’ ” Vo-Doan adds. “Making people feel comfortable with the technology is just as important as making it work.”

Practical Paraborgs

In the future, Vo-Doan and his colleagues aim to address practical issues with the Paraborg. These include compensating for the movements of victims, establishing reliable wireless communications inside collapsed structures, and guiding the cyborgs as they climb over and squeeze through complex environments filled with rubble and dust. Autonomy will become increasingly important for these insects, particularly if the scientists want to operate multiple cyborgs at the same time, he says.

In addition, a fundamental challenge when it comes to working with cyborg insects is that they are living creatures with minds of their own. “We are not piloting them like conventional wheeled robots,” Vo-Doan says. “Electrical stimulation influences their direction, but the insect still generates and controls much of its own locomotion.” To deal with this unreliability, the Paraborgs will need better onboard systems to pinpoint their positions and monitor their actions so researchers can recognize when an insect has deviated from its course or become less responsive.

“We are not suggesting that this is a medical device ready to be used on people today,” Vo-Doan says. “Real disaster sites are full of unstable debris, narrow gaps, and communication difficulties, so we need to understand how the insect, electronics, and injector all perform under those conditions. There are also important questions around drug choice and dosage, sterility, needle safety, reliability, and regulation.”

Ultimately, such research into cyborg cockroaches may help inform robot design, explains Vo-Doan. These insects “can give us useful capabilities sooner while, at the same time, helping us develop the fully artificial systems of the future.”

The scientists detailed their findings last month in the journal Advanced Science.

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.

Humanoids Summit Seoul: 22–23 September 2026, SEOUL

IROS 2026: 27 September–1 October 2026, PITTSBURGH

CoRL 2026: 9–12 November 2026, AUSTIN

Enjoy today’s videos!

I know these videos from Agility can be a little bit silly, but the couch drag in this one is impressive.

[ Agility Robotics ]

Stabilizing unsecured payloads against the inherent oscillations of dynamic bipedal locomotion remains a critical engineering bottleneck for humanoids in unstructured environments. To solve this, we introduce ReST-RL, a hierarchical reinforcement-learning architecture that explicitly decouples locomotion from payload stabilization. Successfully deployed on the Unitree G1 humanoid hardware, this modular approach demonstrates highly reliable zero-shot sim-to-real generalization across various objects and external force disturbances.

[ SteadyTray ]

Thanks, Ioana!

Figure is scaling compute so that its robots can...uh...have their compute scaled, I guess?

Solving for a robot in every home is not a data-and-compute problem, it’s a safety-and-cost problem.

[ Figure ]

The most important thing to know about this gripper is that koalas have two thumbs on each hand.

[ RAI Institute ]

DARPA Triage Challenge Finals are in November!

[ DARPA ]

Online, humanoid robots are very impressive to watch, but behind the scenes, most of those movements are carefully choreographed. Researchers in Carnegie Mellon University’s Safe AI Lab are instead teaching robots how to adapt. Their system, called APEX, allows a humanoid robot to navigate obstacles using adaptive, full-body maneuvers.

[ CMU ]

Researchers from North Carolina State University have created teardrop-shaped soft robots that leap upward or forward when exposed to infrared light—and will keep jumping as long as the light is present.

The robots are made of a liquid-crystal elastomer ribbon shaped like a teardrop, with a thin aluminum tube shaped like a V at one end. When exposed to light from an infrared lamp, the surface of the ribbon contracts, causing the ribbon to rotate. The stiff V at one end of the robot prevents the ribbon from simply rolling in place, causing the ribbon to twist tighter and tighter. This stores energy until the twist reaches a critical point when the ribbon releases that energy, causing the V at one end of the teardrop to snap downward and strike the surface. This launches the teardrop into the air.

[ NC State ]

Thanks, Ship!

I’ll be honest—I was prepared to be underwhelmed by the DARPA Lift Challenge, but there was such creativity in the heavy-lift drone designs that I’m excited for it to come back in 2028.

[ DARPA Lift Challenge ]

Thank you, Christian, for attempting to talk some sense into the internet.

[ Christian Hubicki ]

Humans use not only muscle signals but also stretched skin around joints as a cue for proprioception. To mimic this biological mechanism, we developed a three-layer joint-covering skin with 44 pressure- and stretch-sensitive elements for the musculoskeletal humanoid Musashi-W.

It’s not a replicant, but one day, it will be.

[ University of Tokyo ]

Thanks, Akihiro!

Having mobility issues with your robot? Just staple it to the end of an industrial robotic arm. Problem solved!

[ LimX Dynamics ]

But what if I am the sort of person who needs to speak to a manager?

[ Sharpa ]

In Turpan, China—known as the City of Fire—summer ground temperatures can exceed 50 °C. During the grape harvest, farmers traditionally carry heavy baskets back and forth under the intense heat, while every extra minute in the sun can affect the freshness of the fruit. This year, the DEEP Robotics Lynx M20S joined the harvest.

[ DEEP Robotics ]

This video showcases the achievements of the first OH! GYM! Project cohort, a group of university and graduate students who explored, developed, and deployed their own humanoid behaviors using the open-source AI Sapiens K1 platform. Over the course of one month, the students experienced the complete process of humanoid development—from creating motions in simulation to transferring them onto a physical robot through repeated Sim2Real experiments.

[ ROBOTIS ]

The round is being raised just months after the robot data startup exited from stealth.

Scientific research often depends on complex laboratory equipment that only specialists know how to use. But Anthropic is now rolling out a system that allows AI agents to control lab devices and autonomously carry out experiments.

Laboratory automation technology has been around for decades but getting different bits of equipment to talk to each other has traditionally been a major headache. Most instruments use their own proprietary interfaces, so connecting a microscope to a robotic arm or a liquid handler typically requires bespoke software that takes specialists weeks or even months to build.

Anthropic says its new Model Hardware Standard can reduce this process to minutes by giving devices a common language. It relies on a standardized “driver” that lets any programmable device describe itself to an AI agent, allowing the AI to handle the integration. The company announced it’s opening the system up as a research preview to an initial group of labs and manufacturers.

“Our hope is that the standard can be of use to researchers, engineers, and other practitioners in speeding up the process of discovery and experimentation in any domain that uses devices with a programmable interface,” Anthropic said in a press release.

The standard is similar to Anthropic’s Model Context Protocol, which makes it easier for AI to interact with third-party software, but the new system is aimed at hardware instead. The driver at its heart is essentially a piece of software that sits between a computer and a piece of hardware, translating instructions from one into signals the other can act on.

Most laboratory instruments already run some form of driver, but each has traditionally spoken its own dialect, which is why connecting them has required custom code that can translate between devices. Anthropic’s new driver standardizes that dialect using deliberately simple commands such as “read” or “write,” which can refer to anything from checking a temperature to setting the length of an operation.

Because every device speaks in these same basic terms, machines can find each other on a network and exchange data without a custom program to translate between them. The driver also makes it easier for the company’s Claude agents to learn how to use a device they’ve never seen before.

The standard lets users encode key details, like the weight of a robotic arm, using natural language.  They can either write out their hardware setup themselves or have an agent interview them about it. The system then turns that information into a reference file covering what a device can measure, what can be adjusted, and what safety limits apply.

Anthropic says this lets its agents orchestrate complex experimental processes across multiple instruments in often highly complicated and interactive ways. “We’ve found that Claude interacts with experiments and hardware in an exploratory manner, much as a scientist would,” the company writes. “We observed Claude make an adjustment to a laser, observe the results through a camera to assess how its adjustment moved the laser beam, and repeat the process, seeking to understand the sequence of events.”

Speaking to the Financial Times, Anthropic scientist Alek Kemeny described watching Claude locate a specific, unfamiliar structure in a live brain tissue sample during a neuroscience experiment by manipulating a microscope’s mirrors and lasers on its own. “The neuroscientist sitting there said: ‘Yep, that’s right,’” said Kemeny.

The new standard could be key to the company’s ambition to move beyond its key markets of software development and knowledge work and allow its AI to start having an impact in the physical world. But allowing AI, which is still not immune to hallucinations, to control real-world hardware carries considerable risks.

“It is an impressive proof of concept, but how do we ensure safety in the physical world? Because small errors can matter here,” Kaoutar El Maghraoui, principal research scientist at IBM, said on the company’s Mixture of Experts podcast.

That’s probably why Anthropic is only releasing the standard to a small number of partners initially, and it has committed to working with them to build safety evaluations for AI systems that are operating physical hardware. If the early launch goes well though, AI agents could soon make an impact in far greater swathes of the economy.

The post Anthropic’s Claude Can Now Autonomously Run Science Experiments With Lab Equipment appeared first on SingularityHub.

Less than 24 hours left to apply to host a Side Event during TechCrunch Disrupt 2026 and make your mark in the Silicon Valley scene. Apply before the application closes tonight at midnight PT.

On a good day, the rock quarry in central Texas is about 370,000 kilometers (230,000 miles) from the moon. But last February, when Rishi Jangale watched his 1.8-meter-wide, 150-kilogram inflatable robot roll effortlessly over rocks, gravel, and wet clay, his imagination turned the quarry into the lunar surface instead.

Jangale is an upbeat mechanical engineer nearing the end of his Ph.D. at Texas A&M University, in College Station, Texas. He and his labmates have been working on this big tan “RoboBall” for about five years. Their goal: create a vehicle capable of exploring some of the most inaccessible terrain in our solar system, such as the 21 km-wide Shackleton crater on the moon’s south pole.

The Shackleton crater is 4 km deep and contains many smaller, deeper craters within. Some parts of the crater never see the sun, and within these perpetually dark, frigid pockets lie mysterious substances that planetary scientists have long struggled to examine, including layers of ancient lunar geology and stores of frozen water that could potentially support a future lunar base.

“The moon is like an archive of what happened to the Earth,” says Sara Russell, a cosmic mineralogist with London’s Natural History Museum who is not involved with Texas A&M’s work. “Robotic collection works brilliantly well, and it’s great to see that being explored more in this context.”

“NASA is not going to let astronauts get anywhere near these craters, because if someone falls in, you’re not going to be able to get them out,” Jangale says. “So we thought, what better shape to roll down a hill than a ball?” In a recent paper published in IEEE Transactions on Field Robotics, Jangale’s team reports on the design of RoboBall, a hypothetical lunar mission, and the results from initial tests in the Texas quarry.

Getting Rolling

RoboBall is the brainchild of Jangale’s advisor, former NASA robotics engineer Robert Ambrose, who first conceived of the design in 2003.

Ambrose figured that a ball could address a pesky mobility risk that robots face on lunar terrain, especially in low gravity: tipping over. An inflatable sphere can’t tip over, and the form factor also insulates its internal components from sharp rocks, dust, and the huge temperature swings from over 93 °C in sunlit spots to minus 240 °C in the shade inside lunar craters. Ambrose imagined a wheeled rover parking at a crater’s edge and releasing a RoboBall to explore its depths. Unlike a small tethered rover, the RoboBall wouldn’t roll its way back up—but with no strings attached, it would have a far greater range to collect geological samples, and then be able to launch them back to the rover outside the crater with small rockets.

A robotic rover would ferry RoboBall across the lunar surface to the edge of a crater.R. Jangale, D. Pravecek, et al.

NASA hasn’t brought lunar samples back to Earth in over 50 years. Some morsels of moon geology find their way to Earth as meteorites, but Russell says these lack the “gold standard” field work—context about where that sample actually came from. Even as NASA reboots its crewed moon missions, many lunar sites remain inaccessible.

In 2022, Ambrose’s lab finished a proof of concept, the 0.5-meter-wide RoboBall II. Creating the full-size RoboBall III then took about 11 months. “We were building these robots really quickly,” Jangale says. “Ambrose really encourages us to use and break these robots.” And designs did go awry. Jangale remembers software errors and a drivetrain that proved too weak to roll over soft bumps in initial tests.

Texas A&M RAD Lab/YouTube

How to Slow Your Roll

RoboBall drives by moving a pendulum within its shell, shifting its entire center of mass. On flat ground, if the pendulum’s arm points forward, the shell rolls forward to compensate, and as long as the pendulum keeps the center of mass in front of the center of the ball, RoboBall will keep rolling forward. If the pendulum leans a few degrees to the left or right, RoboBall steers left or right to match. “The robot wants to go where you point the pendulum,” Jangale says. The 340-pound ball is just soft enough to bounce lightly over bumpy obstacles, but its slight overpressure keeps it relatively firm. On steep slopes, this mechanism also lets the ball control its downhill speed by simply angling the pendulum uphill.

“The beauty here is the simplicity,” says Hiro Ono, an aerospace engineer who worked on robot mobility at NASA’s Jet Propulsion Lab for 13 years before joining Georgia Tech. For space robots, Ono describes simplicity in terms of the number of actuators. RoboBall has just two, and both are fully inside of the shell, shielded from environmental risk factors like dust, a feature that Ono describes as “unique.”

After the first quarry tests, it took about seven months for the team to design and build an upgraded RoboBall III with 2.5 times more torque—enough to fling itself over small obstacles and roll up 20 degree slopes.

A small rocket can launch out of the center of the RoboBall to return a sample back to a rover outside the crater.R. Jangale, D. Pravecek, et al.

Getting Mission Ready

In the new paper, the remotely operated RoboBall III descended the quarry’s slopes, navigated soft terrain, and launched hypothetical payloads back out of the crater with small rockets. Powered by a large battery, the robot would inflate itself during a lunar mission and charge up from a robotic rover at the edge of a crater before heading out on its own.

RoboBall is probably not the right platform for all kinds of missions—its slightly bumbling nature means that it’s not ideal if you want to collect a sample from a very specific rock, for example. For now, the team hopes to work with scientists designing lunar science instruments to physically fit into RoboBall’s carry-on-luggage-size interior, while also fitting in with how RoboBall operates. “The robot is not the mission,” Jangale says. “The robot is a way for you to complete the mission.”

The current version of RoboBall cost roughly $250,000 to build, but is not quite ready for the moon in its current form. While its gold-treated aluminum parts are appropriate for space missions, its other materials are not. Space-grade electronics will cost more, and the ball’s shell—made from a material tough enough to roll over steel shards and withstand minus 184 °C temperatures—has never been evaluated in lunar extremes. Aside from materials questions, the team plans to engineer the ball to adapt how it drives on varying slopes autonomously. They also hope to collaborate with government agencies or spaceflight contractors to keep refining the robot’s design for a real, eventual mission.

Later this year, Texas A&M will open a new facility in Houston with the world’s largest indoor simulated moon and Mars landscapes. The facility is only 190 km from Jangale’s quarry. It’s still about 370,000 km from the moon, but it’s going to help Jangale get his robot quite a bit closer.

The Builders Stage is returning to TechCrunch Disrupt, bringing together founders, startup operators, and investors for practical conversations on what it takes to build and scale.

A perception stack is shaped by the vehicle that carries it. Move the same software to a new carline—for example, from an SUV to a sedan or another vehicle...

A perception stack is shaped by the vehicle that carries it. Move the same software to a new carline—for example, from an SUV to a sedan or another vehicle variant in the portfolio—and its perception of the world changes. The sensor placement, calibration, fields of view, occlusions, body geometry, timing, and coverage all shift. A traffic light may appear in a different part of the frame.

Source

The U.S. is shutting out more foreign-made drones and robots. China’s scale means the global competition may simply move elsewhere.

Anthropic's Model Hardware Standard (MHS) gives AI agents a unified interface to physical devices like robotic arms and lab instruments. In early tests, integration time dropped from weeks to hours. But Claude sometimes failed to grasp physical cause and effect, so human oversight remains essential for now.

The article Anthropic wants to do for physical hardware what its Model Context Protocol did for software appeared first on The Decoder.

💾

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.

Humanoids Summit Seoul: 22–23 September 2026, SEOUL

IROS 2026: 27 September–1 October 2026, PITTSBURGH

CoRL 2026: 9–12 November 2026, AUSTIN, TEXAS

Enjoy today’s videos!

Nvidia just paid US $12.9 billion for the company that acquired Pollen Robotics, and this must be why.

Meet Microduck. 🦆 The 25-centimeter, 780-gram robot that waddles, falls, gets back up, and learns new tricks.

Packed inside: 15 degrees of freedom, a front camera, an 8x8 lidar, two IMUs, mics, a speaker, NFC, Wi-Fi, and Bluetooth.

Out of the box, Microduck already walks, sits, crouches, roller skates, picks up objects with its articulated beak, and recovers from falls on its own. Drive it with a game controller, plug-in accessories, and NFC tagged objects, run autonomous behaviors, or gather several Microducks for races and football.

Software fully open source. Ready for whatever you throw at it.

On pre-order for an astonishingly low $399, and ships before Christmas.

[ Microduck ]

Thanks, Matthieu!

If you’ve chosen to ignore all the earlier DARPA Lift Challenge videos that we’ve posted, now you can get all caught up in about five minutes.

[ DARPA ]

You had me at “54-gram robot that out-jumps a kangaroo.”

[ IEEE Transactions on Robotics ]

Sometimes, you just need a video like this.

Most fish-inspired robots are built for one size and one job, so scaling them up or down usually means starting from scratch. A team of engineers says it has found a way to solve that problem. They’ve unveiled ScaFi, a robot modeled on fish like cod and mackerel.

[ New York University ]

Thanks, Leah!

Martin writes, “We’re a small robotics team in Czechia, Europe, building practical hardware around the Unitree G1. Here’s a short demo of our lightweight gripper picking up a strawberry; the gripper weighs under 200 grams and is designed for simple, sensitive manipulation without adding a complex multifinger hand.

[ Sentio Robotix ]

Thanks, Martin!

Hybrid visual markers that are useful for both cameras and lidar is a neat idea.

[ Hello Robot ]

Thanks, Binit!

EmoLo brings emotion-inspired expressive locomotion to Open Duck Mini V2, a low-cost, open-source bipedal robot inspired by Disney’s BDX droids. With a single reinforcement learning policy, the robot can generate distinct walking styles associated with different emotional expressions, showing how characterful and expressive whole-body motion can be achieved on an accessible robotic platform.

[ EmoLo ]

Thanks, Masato!

If it’s possible for a robot with a completely immobile face to look frustrated, this robot absolutely does, starting at three minutes into this video.

[ DLR RM ]

Noble Machines deployed its first general-purpose robots to a Fortune Global 500 industrial customer within 18 months of the company’s launch and met its first delivery milestone, made possible by its AI-driven whole-body control and industry-leading end-to-end autonomy.

[ Noble Machines ]

We’ve reduced the time it takes to go from physical prompt → robot behavior. The faster anyone can teach a robot to do something new, the easier it becomes to scale physical work.

[ Generalist ]

I know this video is mostly a gimmick, but I would totally rent a moderately heavy lift quadruped for a couple of days to help with a move.

[ DEEP Robotics ]

Is taking two minutes to excellently fold a shirt too long, or do we even care how long it takes, as long as it’s a robot doing it?

[ Tokyo Robotics ]

TRON 2 × Wuji Hand 2 handles TCM pharmacy work: picking, weighing, grinding, and packaging. The omnidirectional base frees the hands, while precise gripping and dual-arm force control enable midair operations.

[ LimX Dynamics ]

Person who genuinely knows things about robots, Christian Hubicki, explains everything about robots smashing into walls.

[ Christian Hubicki ]

Glass breaks. That single fact has made automotive glass manufacturing one of the most stubbornly difficult lines to automate, and one of the most expensive to run manually. A 1.5-meter laminated windshield blank weighs around 20 kilograms, flexes unpredictably under load, and shatters the moment positioning goes wrong. For decades, factory managers accepted high breakage […]

Clem Delangue, CEO of Hugging Face, said the Microduck is an “open-source robot you can teach new tricks with reinforcement learning.”

This article includes an affiliate link Tennibot, which says it is “the only company designing, engineering, and assembling AI-powered ball machines in the United States”, has launched Partner Lite, a new AI-powered ball machine that delivers the company’s signature intelligent training experience at its most accessible price yet. Available for a special introductory price of […]