When robots and humans work together in a factory, they are usually kept clearly separated – in part for safety reasons. The robot takes on a task, completes it, and hands it over to workers, who carry out the next step in the process.

“This can be physically demanding, for example, during quality inspections of components that have to be repeatedly lifted and set down,” says researcher Federico Masiero from the Chair of Intelligent Bio-Robotic Systems at the TUM School of Computation, Information and Technology.

Developed at the Munich Institute of Robotics and Machine Intelligence (TUM MIRMI), led by Prof. Lorenzo Masia, the so-called WearaCob (short for “wearable” and “collaborative”) is designed to make lifting and carrying easier for people with the help of an upper-body exoskeleton, enabling assistance from a single-arm, collaborative robot, or “cobot”.

Exoskeleton reduces upper-arm effort by up to two-thirds

The exoskeleton is worn like a backpack, with an electric motor mounted on the back. On both sides, thin, high-strength wires run from the back over the shoulders to the front, where they are attached to elbow-pad-like supports. When the motor pulls on the wires, they take over part of the work normally performed by the biceps muscle in the upper arm. A single-arm robot assists the person wearing the exoskeleton. When the robot picks up and hands over an object, it weighs it and wirelessly transmits its weight to the exoskeleton. The cables, which in their default setting precisely compensate for the weight of the wearer’s arms, then provide additional assistance proportional to the weight of the object.

Recent studies show that this configuration can reduce muscular effort by up to 65%. “To handle asymmetrical components as well, the robot also determines their center of mass,” explains researcher Masiero. “The advantage is that one arm can receive more support than the other, making it possible to balance uneven loads.”

The shoulder exoskeleton also works well on its own, but a bit differently: The most common research method is to measure muscle activity in the upper arm and use this to determine how much additional assistance is required at that moment. This method is fairly accurate, with an error margin of only 0.5 to 1 kilogram. However, the sensors must be attached to the upper arm before each use of the exoskeleton, which is rather impractical in a factory.

Cobot programming by demonstration

The single-arm robot is a typical seven-joint cobot. This makes it mobile and flexible, while its ability to automatically slow down as it approaches people makes it safe enough for use in a research setting.

The combination of the exoskeleton and the robotic arm offers particular advantages for industry: “We were able to demonstrate not only that we can specifically ease the physical workload of factory workers, but also how easy it is to teach the cobot new tasks,” explains Prof. Masia. “We can program it by guiding the robotic arm. Not a single line of code is needed. That’s a huge advantage over many robots currently used in factories, which operate behind safety barriers and away from human workers.”

For the past few weeks, the research team has had a new cobot from NEURA Robotics in the laboratory, where it is set to learn additional skills in the TUM RoboGym in the near future. Not only can it handle components weighing up to 15 kg – five times more than its predecessor – but it is also equipped with a microphone, an integrated 3D camera and two other cameras. Researchers can give the latest cobot instructions directly, and it “sees” the person interacting with it in 3D as a human avatar. Special thanks go to the project partners Akina, a Swiss startup that uses computer-vision algorithms and webcams to track users’ movements, and Neura Robotics for providing state-of-the-art robotics technology.

The Munich Institute of Robotics and Machine Intelligence (TUM MIRMI) is an integrated research institute at the Technical University of Munich (TUM) focused on robotics and AI.

ROBEAR was 150 cm tall, weighed 140 kilograms, and had a cute bear face. Developed in 2015, the care robot had been designed for lifting nursing home residents, but never made it past the prototype phase. The more compact lifting robot “Hug” also failed tests in Japanese nursing homes. Both products were time-consuming, cumbersome to manoeuvre, and the lifting process was uncomfortable for many residents.

Questioning the roles of technology

ROBEAR and Hug are representative examples of unfulfilled expectations placed on robots and artificial intelligence against the backdrop of aging populations and the shortage of care workers. “The fact that they haven’t been successful in practice yet is largely due to the technology design’s falling massively short of the complexity of caregiving activities,” says Laura Vogel, a doctoral candidate in the Department of Industrial Engineering and Organization at TU Wien. “The notion that one can simply install a device that will solve problems is unrealistic.”

Vogel’s colleague Reinhard Kletter, a doctoral candidate at the Institute of Management Sciences at TU Wien, is a robotics expert and has seen quite a number of care robot prototypes come and go: “Technology often looks amazing on paper. But if it doesn’t meet the daily needs and values of care, it’s a waste of time and effort.” And Vogel emphasizes: “With many solutions, the effort required for implementation and maintenance is currently underestimated. That requires frequent software updates that simply shift costs and effort to the IT sphere instead of saving them altogether.”

User-centred research, on the other hand, seeks to focus product development on the needs of care facilities. Designs often start with a specific technology, with the objective of applying it to the care context.

Participation and dialogue

Caring Robots // Robots in Care is a transdisciplinary project funded by the Austrian Science Fund FWF in which Laura Vogel, Reinhard Kletter, and other researchers from the fields of computer science, robotics, and social sciences are taking a step back: “Together with care workers and care recipients, we defined what roles for robots and artificial intelligence are desirable and useful in the care context,” notes Kletter.

In order to understand the needs and challenges in care settings, Vogel and other social scientists interviewed caregivers and systematically observed their daily work routines. In addition, the project team organized a series of workshops with caregivers and residents in Caritas nursing homes. “The goal was to learn from one another. First, we presented the technical background of AI and robotics: how do robots ‘see’, how can one communicate with machines via language models, and what possibilities arise from that?” explains Kletter.

Finally, the researchers, together with the workshop participants, collected wishes and ideas for applications in the care context. Building on this basis, Kletter and his colleagues from the fields of computer science, electrical engineering and robotics are currently developing several prototypes.

Language models instead of robots?

“We found that for many desired instances of use one doesn’t really need a robot with a physical body,” Kletter concludes. Caregivers are particularly hoping for support with documentation duties. Kletter has therefore developed AI-powered assistance software that is currently being tested and continuously improved. The idea: caregivers wear a small clip-on microphone that records conversations during care activities.

“This is important because even small talk can contain care-related information,” says the researcher. By means of speech recognition and large language models, care-relevant information is subsequently extracted and used to generate a structured report that can be used for care documentation. In addition, the researchers are testing another prototype to see how language models can be usefully applied to facilitate conversation and biographical work, particularly for people with dementia.

Value-oriented technology

It is not enough for care technology to be functional. It must also be aligned with the ethics of care. In his doctoral thesis, Reinhard Kletter focused on the issue of privacy, to which residents and caregivers attach special importance: “Privacy must be considered from the very beginning of the design process. This is not just about data protection, but also about ensuring that a tool does not transform the interaction between caregiver and resident.”

In the case of the AI documentation tool, only the transcript containing care-relevant information is retained, whilst the audio file is deleted immediately. The mere possibility that the technology could also be used for surveillance is a problem and requires clear communication and transparency.

Care quality first, hype second

The privacy aspect shows how new technology can change the social fabric in a work context, says Laura Vogel. She is currently focusing particularly on the effects on the professional self-image of nursing staff: “Technology should create added value, i.e. make work easier or strengthen skills, rather than downgrade nurses. That might be the case, for instance, if they have to clear away a machine after use and thus ultimately have less time to spend on care.” Kletter points out that the autonomy and participation of the caregiver must be preserved in all care processes. Therefore he wants to incorporate “human-in-the-loop” steps into AI-supported care documentation.

Many of the wishes expressed in the discussions – such as for a robot that can do cooking together with residents – reflect a fundamental difficulty, according to Vogel: “Care should be time-efficient, but ideally it should also be holistic and person-centred. I don’t think these conflicting objectives will be resolved if machines break the work down into many small, automated steps.”

When considering what care should look like in the future, we as a society should critically examine the advantages and disadvantages of AI and robotics – without succumbing to the hype and blind faith in technology.

Robots in space exploration can help with everything from resource extraction on the Moon to the maintenance of space stations. But that requires the robots to be extremely robust. Graphic illustration: ESA

An ESA-funded project will develop a new protective covering, intended to pave the way for more affordable robots in space, while also holding potential for terrestrial applications.

Future space exploration will increasingly rely on robots as the primary workforce. However, this requires them to become better equipped to operate in the extreme environments of the Moon, Mars, and in orbit – with abrasive dust, intense solar radiation and temperatures ranging from minus 150°C to plus 120°C.

The European Space Agency (ESA) has now appointed a pan-European consortium led by Danish Technological Institute (DTI) to develop the next generation of a protective cover for robotic arms. The project is called Smart Skin for Exploration Cobots and aims to advance the technology to a level where it can be demonstrated under space-like conditions.

“The potential for robots in space exploration is extensive. They can help with everything from resource extraction on the Moon to on-orbit satellite servicing and active debris removal. But this requires the robots to be extremely robust and capable of operating autonomously – or safely in collaboration with humans,” says Christian Dalsgaard, Senior Consultant at DTI.

Advanced multilayer protection

The smart skin technology is being designed to be adaptable to different robotic arms – both for upcoming lunar missions, future Martian missions and for in-orbit operations.

At its core is a 3D-printed scaffold that can be mounted on the robotic arm. It serves as a platform for four integrated functions: a thermal and dust-protective layer that shields against extreme temperature fluctuations and abrasive dust penetration; flexible power and data cabling; sensors capable of detecting and preventing collisions; and features that enhance human-machine interaction.

3D printing has been chosen because it offers the necessary design freedom, but the technology will be pushed beyond its comfort zone – with entirely new approaches to design and material selection.

Traditionally, Multi-Layer Insulation (MLI) materials have been used on all spacecrafts, providing high-efficiency thermal protection for the whole structure – or just for smaller instruments. However, these applications are static without any motion. Developing a similar type of thermal insulation for moving parts is significantly more challenging, but it allows for a wide range of future applications for robotic systems.

“Applying an advanced protection system could lead to building robotic arms from commercially available components. This can create a cost-effective way of providing new solutions for customers in many space domains – from deep space missions, through in-orbit servicing to Moon colonisation. At Admatis, we are committed to any development that gives Europe a competitive advantage, and this project is fully in line with our strategy,” says Tamás Bárczy, CEO at Admatis.

From space to practical benefits on Earth

Although the smart skin technology is being developed specifically for the unique challenges of space, parts of the technology may in time have wider application potential.

“We see strong potential for the technology eventually to find applications in companies where robots are exposed to extreme conditions. Think of metal foundries, where dirt and extreme heat challenge equipment performance. The technology we are developing could potentially extend the service life of critical equipment and reduce maintenance costs,” explains Christian Dalsgaard.

International collaboration with specialist expertise

The project builds on a previously successful pilot phase and brings together leading European space companies and specialists within adjacent fields.

DTI is coordinating the activities and contributing specialists in robotics, functional materials science and industrial 3D printing.

Admatis (Hungary) is developing the thermal protection, while PIAP Space (Poland) and Redwire Space Europe (Luxembourg) are making their expertise and robotic arms available – the same arms currently being developed for ESA’s upcoming lunar missions. This ensures that the smart skin technology is designed from the outset for the specific systems it is intended to protect.

As part of the ForNeRo research project, Professor Dirk Wilhelm (right) and researcher Luca Wegener (left) are working together. Image courtesy of Astrid Eckert / TUM

How can robots and humans work together as effectively as possible in the operating room of the future? Researchers from the Technical University of Munich (TUM) and TUM University Hospital investigated this question as part of the ForNeRo research project. Using a sensor-equipped system, they analysed surgeons’ movements during procedures and collected data from simulated robot-assisted operations.

Five depth cameras mounted on the ceiling of the experimental operating theatre at TUM University Hospital in Munich generate a three-dimensional digital image of the room 15 times per second –a digital twin of the surgical environment. At the operating table stands Prof. Dirk Wilhelm, Head of the Chair of Medical Robotics at TUM and a surgeon and senior physician at TUM University Hospital. He is wearing a suit fitted with motion markers on the joints and head, tracked by an infrared system with ten cameras. Microphones record and spatially locate conversations within the surgical team, while additional physiological data is collected to measure stress levels among staff.

The aim of this sensor data and the digital twin is to improve surgical workflows, integrate robotic assistance systems into clinical workflows as efficiently and ergonomically as possible, and ultimately reduce the workload of medical staff. The sensor system developed by Prof. Wilhelm’s research group for minimally invasive interdisciplinary therapeutic intervention (MITI) is now being used for the first time in Germany to collect data from a real operating room environment. “In the next step, this data could help improve the use of robots in surgery,” says Prof. Wilhelm. All data collection in the operating room requires the consent of patients and all parties involved.

Testing robotic systems in routine surgical procedures

For robotic systems to assist in future operating rooms, researchers will need more than sensor data alone. The experimental operating theatre therefore serves not only as a data collection platform, but also as a test environment for robot-assisted procedures on anatomical models – in other words, testing collaboration with robotic assistants.

As part of the ForNeRo research project, the researchers investigated three common minimally invasive procedures: gallbladder surgery, inguinal hernia repair, and sigmoid resection, the partial removal of the large intestine. Two robotic systems were used in each procedure. The first, Solo Assist II, held and positioned the endoscope. The second was MIRO, a modular surgical robot developed by the German Aerospace Center (DLR) that surgeons can control using a joystick and other interfaces. During a simulated procedure, the surgeons used MIRO to manipulate a miniature gripper, position a plastic mesh during hernia repair and assist with suturing.

A robot can carry out simple tasks

The simulated operations are designed to help configure robotic systems that can assist surgeons during minimally invasive procedures. To evaluate their potential, Max Bergholz from the Chair of Ergonomics at TUM records surgeons’ postures and movements in a sensor-equipped operating room while performing procedures on anatomical models. Participants are also asked to assess the physical and mental strain experienced during the different phases of the operation.

“Surgeons often report back pain caused by maintaining rigid postures for long periods,” says Bergholz. “Earlier systems also required them to operate as though looking into a mirror, forcing them to constantly adapt their spatial orientation.” His goal is to make surgical work as ergonomic and intuitive as possible.

Robotic systems eliminate much of this need for readjustment. They allow surgeons to operate with greater precision, since larger hand movements with the joystick translate into movements of only a few millimetres inside the body. Unlike established systems like the da Vinci Surgical System, the new system also allows the surgeon to remain physically closer to the patient.

The research showed that robotic assistants can already take over simple tasks in an operating room – such as holding an endoscope – without increasing surgeons’ workload. “This allows us to explore how robotic assistants can be seamlessly integrated into clinical workflows,” says Bergholz.

AI is expected to better understand surgical procedures

Looking ahead, TUM Professor Dirk Wilhelm sees potential for using the complex data from the operating theatre for artificial intelligence applications. “Data are a fundamental building block for AI systems in the operating room,” says Wilhelm. “Such systems could automatically recognize which surgical instruments are being used and identify the organs being operated on.”

The initial goal is to improve surgical workflows and planning processes. In the longer term, AI could help decide when a robotic assistant would be beneficial.

Annual shipments are projected to approach 1.8 million units by 2036, driven primarily by automotive manufacturing, with logistics following and home-use remaining a longer-term opportunity with limited penetration within the forecast period.

This growth is supported by the accelerating push toward Industry 5.0, rapid progress in embodied AI, continuous improvements in materials and component supply chains, and sustained strategic backing from investors and OEMs. Compared with open or highly unstructured environments, industrial settings such as automotive manufacturing offer more standardized workflows, clearer task boundaries, and stronger labour-cost pressure. These conditions make them more likely to become the first scalable deployment markets for humanoid robots.

At the same time, declining hardware costs are reshaping the economic baseline. IDTechEx analysis indicates that the average selling price of humanoid robots is expected to fall from approximately US$114,700 in 2024 to around US$37,000 by 2030, with further reductions expected into the mid-2030s. As capital costs decline, the cost per productive hour falls accordingly, with the most significant reductions occurring during the early stages of commercialization. However, while cost reduction is a necessary condition for adoption, this alone is not a sufficient reason to adopt. The IDTechEx report,”Humanoid Robots: Market, Technologies, and Opportunities 2026-2036″, provides a detailed analysis of humanoid robot market forecasts, cost evolution, ROI scenarios, technology readiness, and key application opportunities.

Comparison with human labour remains scenario-dependent

Based on IDTechEx interviews with major industry participants and amortization calculations across representative commercial deployment scenarios over the next decade, the operating cost of humanoid robots is expected to remain highly dependent on deployment efficiency. Unlike fixed automation systems, humanoid robot utilization can vary significantly by task type, workflow structure, environmental complexity, and system integration level. IDTechEx therefore incorporates multiple utilization scenarios in its modelling to reflect a range of real-world deployment conditions, including high-, medium-, and low-efficiency cases.

IDTechEx’s scenario-based modelling suggests that humanoid robot operating costs could vary significantly depending on deployment efficiency. At the current early-commercialization stage, costs remain highly sensitive to utilization, task continuity, and integration quality. However, as enterprise procurement prices decline and deployment experience improves, high-utilisation industrial scenarios could bring operating costs below US$5/hour by around 2030, with further reductions possible toward 2036.

At face value, this cost level is increasingly attractive when compared with human labour costs. In high-labour-cost markets such as the US, total employer cost is expected to continue rising steadily. In China, labour cost starts from a lower base but grows at a faster rate, reinforcing the long-term economic rationale for automation. However, this comparison needs to be interpreted carefully. A robot’s cost per hour is not directly equivalent to a human labour cost per hour, as it depends on sufficient utilisation, task continuity, and operational stability, all of which remain variable in current deployments.

As a result, humanoid robots are beginning to show cost competitiveness, particularly in high-utilisation industrial scenarios. However, in medium- or low-utilisation settings, the cost advantage can be significantly reduced even as hardware prices fall. In other words, the cost curve is improving, but whether the cost advantage can be realized depends strongly on the deployment environment.

Profitability depends on effective output

From an ROI perspective, IDTechEx calculations suggest that humanoid robots are beginning to show a clear payback pathway under favourable deployment conditions. By 2026, payback periods can be reduced to around 6 months under high-utilisation scenarios, compared with approximately 15 months under medium utilization. As hardware prices continue to decline and deployment experience improves, ROI feasibility is expected to strengthen across a broader range of industrial applications.

However, a shorter payback period should not be interpreted as guaranteed profitability. The core variable in humanoid robot economics is not only equipment cost, but the effective value of the work delivered by the robot. In practical terms, this means whether the robot can perform economically valuable tasks consistently, reliably, and at a sufficient level of productivity across different environments.

This remains the main bottleneck for large-scale adoption. Humanoid robots are becoming increasingly feasible in selected structured industrial environments, but capability limitations remain clear in complex, variable, or safety-critical tasks. In the near term, more realistic deployment pathways are likely to prioritize high-labour-intensity, repetitive, standardized, or hazardous tasks where the economic case is easier to validate.

Overall, IDTechEx believes that the cost advantage of humanoid robots is becoming increasingly visible, and ROI can already be demonstrated in selected deployment scenarios. However, large-scale commercialisation will depend on continued improvements in software capability, task generalisation, system integration, and deployment efficiency, rather than hardware cost decline alone.

The new robot from Prof. Angela Schoellig’s TUM Learning Systems and Robotics Lab looks like a broomstick on wheels with a camera mounted at the top. It is one of the first robots that not only integrates image understanding but also applies it to a clearly defined task.

To find a pair of glasses misplaced in the kitchen, for example, the robot has to look around and build a three-dimensional image of the room. The camera initially provides two-dimensional images, but these pixels also contain depth information. This creates a spatial map of the environment that is accurate to the centimetre and is constantly updated. A laptop also provides the robot with information about which objects are visible in the image and what significance they have for humans.

“We have taught the robot to understand its surroundings,” says Prof. Angela Schoellig. The head of the Robotics Lab at the TUM Chair of Safety, Performance and Reliability for Learning Systems aims to develop robots that can navigate any environment independently. Humanoid robots working in factories or robots in care settings in private homes require this newly developed basic understanding, which, as Schoellig explains, “is important for all robots that move in spaces that are constantly changing”.

Internet knowledge translated into the robot’s language

The robot therefore understands that a table or window sill can be used to briefly set down a pair of glasses, whereas a stovetop or a sink are not suitable for this purpose. “The language model captures the relationships between the objects and we convert this information into the robot’s language,” explains Prof. Schoellig.

Two-digit numbers appear on the three-dimensional map of the environment, constantly recalculating the likelihood that the object being searched for is located there. According to the research results, the robot then searches the probable locations almost 30 per cent more efficiently than if it searched randomly throughout the room. Artificial intelligence is used in two ways: on the one hand in image recognition and on the other hand through the use of a language model.

Another special capability of the robot is that it remembers previous images and is able to compare them with new images of its surroundings. If a new object suddenly appears in the kitchen, it recognizes the change with a high degree of certainty (95 per cent) and marks these areas as “highly probable” search locations.

Next step: searching behind cupboard doors

In the next step, the TUM scientist and board member at the Munich Institute of Robotics and Machine Intelligence (TUM MIRMI) also wants to search for objects that are in a drawer or behind a door. To do this, however, the robot will not merely have to draw on knowledge from the internet but will also have to interact with its surroundings.

Robotic arms and hands must open a cupboard and determine whether it opens upwards or sideways and how best to grasp the handle. This will enable the robot to search even in closed spaces such as cupboards or drawers.

The Garmi has evolved: on a stable, mobile base, the upper body of the new care-assist robot is attached to an extendable lifting column with arms located on the right and left-hand sides of the column. Above the arms is a head with alert eyes that blink from time to time. The new generation of Garmi has various sensors: cameras are mounted at eye level to detect movements in the environment, a lidar at leg height keeps objects in the immediate vicinity at a sufficient distance, and in future, 3D cameras will secure and coordinate the movements of the two arms. There is also a screen at chest height.

“The new Garmi understands language, develops a plan independently and brings a patient something to drink,” says Alexander König, whose team developed and implemented the new platform. Based on the new design, the first forward-looking functionalities have now been developed. The MIRMI professor says: “A robot must be functional and operable, but must also have an appealing appearance. That’s why we are collaborating with design experts.”

Robotics engineer König sees his Garmi research team as an integrator that brings technology and design together. This includes, for example, precise grasping functionality (perception) and the ability to arrive at the exact location where a task is to be performed (navigation). The design should also support people in interacting and communicating with the robot while conveying trust and safety.

Bavaria’s health minister Judith Gerlach said: “The new development approach of Garmi is extremely exciting. At its research site in Garmisch-Partenkirchen, TUM is creating innovative solutions that are ideally suited to relieving the burden on nursing staff. At the same time, the quality of life of people in need of care can be improved.”

TUM Vice President Gerhard Kramer adds: “The geriatronics research team in Garmisch has once again demonstrated that it is closely attuned to the needs of caregivers and older people. This is the only way to find solutions that ultimately provide optimal support for those in need of care. It’s great to see that the new Garmi was developed in such close collaboration with the Munich Design Institute.”

Needs of caregivers, patients and clinicians

“While the original Garmi was designed as a versatile research platform, the new Garmi has been specifically developed for the care context,” says Annette Diefenthaler, Professor of Design and Transdisciplinarity and Director of the Munich Design Institute (MDI), who worked with an external partner to develop the design of the new robot. Several workshops attended by carers, elderly people, doctors and robotics researchers provided important insights for the design of the new generation of assistance robots.

‘Empathetic, competent, professional, trustworthy and friendly’ were among the characteristics listed on the wish list of participants at a design workshop held at the end of last year. Despite the technology involved, it was clear that acceptance and emotional closeness to a care robot play a decisive role. This is one of the reasons why it is clad in loden, a traditional wool fabric common in the Alpine region: “The fabric combines tradition and the future, gives the robot warmth and regional character and makes it more trustworthy,” comments Prof. Diefenthaler.

A friendly being that controls technology

The robot is more like a mobile platform than a humanoid. “But it was clear to us from the outset that it should come across as friendly and approachable – with subtle human-like features,” says Diefenthaler. The new platform does not look like a human being: “It’s a friendly creature that controls technology. This allows the machine to fade into the background while the robot creates an emotional connection.”

The new Garmi can pick up objects from the floor, but also retrieve them from up high. Unlike the first-generation Garmi, the face and screen are separate. In future, when a doctor is connected for a remote examination, their head will appear on the screen, just like in a video call. The next step is to make the new Garmi safe for use in both care facilities and the home environment of senior citizens. Bringing a drink to a thirsty person is only the first step. “Helping people get up, enabling communication and participation in social life, reminding them to take their medication – the possible applications are wide-ranging,” says Prof. König.

Data collected using the gloves – which cost around £50 to make – could be used to teach robots to use their hands in similar ways to humans, the team says.

Developing robots with greater dexterity could improve their ability to be used in challenging applications, such as remote surgery, virtual reality and carrying out tasks in space.

Each glove is equipped with a range of sensors that can detect subtle movements, such as finger bending and changes in the spacing between fingers – a feature that similar, existing technologies tend to lack.

The sensors, housed in silicone and composed of electrodes made of liquid metal, detect movement by measuring changes in the amount of electrical charge – known as capacitance – stored by their electrodes. Changes to capacitance are produced when the fingers of the glove bend or the distance between them changes.

Researchers from the University of Edinburgh tested their design by collecting hand gesture data from six participants. While wearing the glove, each participant performed 30 different hand gestures, which the sensors detected with more than 99 per cent accuracy.

To investigate the sensors’ ability to track even more complex hand motion, researchers tasked participants with performing random movements whilst wearing the glove.

The team used cameras to track the hand at the same time, producing a comparison dataset to test the glove’s accuracy. Their results show that the sensors can accurately reconstruct hand shape and movements that closely match the comparison data, outperforming current technologies by almost 10 per cent.

Following on from this study, researchers are seeking to improve the glove’s sensing capabilities by integrating technology to mimic the human hand’s sense of touch across the whole palm.

The research was presented at the 2025 IEEE/RSJ International Conference of Intelligent Robots and Systems in Hangzhou, China – one of the largest and most impacting robotics research conferences worldwide. The work was supported by the European Research Council, which has also recently awarded the research team a Proof of Concept grant to commercialise its related, flexible electronic skin technology.

The team is working with Edinburgh Innovations, the University’s commercialisation service, to translate these proprietary technologies into real-world impact. The focus is on next-generation robotics, particularly dexterous and humanoid robots, where rich whole-body sensing is critical for safe, intelligent interaction with the physical world. Target applications also span a range of use cases from health care, such as surgical robotics and compliant protheses, to virtual and augmented reality and wearable technologies.

Dr Yunjie Yang, of the University of Edinburgh’s School of Engineering, who led the study, said: “By using highly stretchable liquid metal electrodes, we can capture the continuous, fluid transition of a hand in motion. This high-fidelity gesture data is the missing link needed to teach robots not just how to hold an object, but how to manipulate it with human-like agility and grace.”

Through a dual-city setup connecting the Beijing Humanoid Robotics Data Training Center and the IROS exhibition booth in Hangzhou, a RealMan trainer in Beijing remotely controlled humanoid robots at the booth to perform complex interactive tasks such as handing over a towel and passing fruit.

Accelerating embodied intelligence innovation

RealBOT is a comprehensive open platform designed to accelerate embodied AI innovation. It integrates advanced motion control, multi-dimensional perception, and precision manipulation, empowering researchers and developers to experiment, prototype and advance embodied AI applications.

Built for high-quality data collection, RealBOT leverages over one million multimodal data samples collected across ten real-world application scenarios from RealMan’s Data Training Center. This enables more robust model training and faster deployment across industries. Key technical advantages include:

• Full-stack in-house development: Proprietary actuator and control technologies ensure optimised performance and reliability.
• Flexible compute support: Compatible with both NVIDIA Jetson Orin and Digua RDK S100 platforms.
• Multisensor fusion perception: Integrates depth and wide-angle cameras, LiDAR, IMU, and microphone arrays.
• Compact design for narrow spaces: Excellent mobility and precision operation in confined environments.
• High-degree dexterity: 21 active DOFs with support for dexterous hands and adaptive grippers.
• Open ecosystem: Compatible with various mainstream vision systems and gripper models.

With its ultra-lightweight architecture, RealBOT lays the groundwork for scalable humanoid systems and high-quality data generation, empowering embodied AI to move from lab research into homes, factories, and service industries worldwide.

Driving global collaboration in embodied AI

The Beijing Humanoid Robotics Data Training Center, built by RealMan, now serves as the backbone for high-quality data generation and AI training. Featuring over 100 robots across 10 real-world environments, it enables the continuous collection of large-scale multimodal datasets that support RealBOT and global research partners, building an open, collaborative ecosystem for embodied intelligence.

Image courtesy of Tara Winstead from Pexels © 2021

Humanoid robots are moving closer to real-world deployment – and their progress depends on physical intelligence and real-time reasoning. With the recent announcement of general availability of NVIDIA Jetson Thor, Analog Devices (ADI) is further accelerating the development of humanoids and autonomous mobile robots (AMRs).

Combining ADI’s edge sensing, precision motion control, power integrity and deterministic connectivity with Jetson Thor’s compute capabilities, Holoscan Sensor Bridge and Isaac Sim, creates a path to scale reasoning-enabled robots from simulation to deployment.

Jetson Thor redefines what’s possible for robotics. With a NVIDIA Blackwell GPU, transformer engine, Multi-Instance GPU (MIG), a 14-core Arm Neoverse V3AE CPU, and up to 128GB of LPDDR5X memory, it delivers 2070 FP4 TFLOPS server-class AI compute in a mobile power envelope. Its high-throughput I/O, including 4×25GbE, provides the bandwidth needed to fuse dense multimodal sensing in real time.

This capability makes NVIDIA Jetson Thor the first platform to run robotics foundation models at scale, from vision-language to vision-language-action models, enabling robots to move beyond perception into reasoning and physically intelligent behaviour. That aligns directly with ADI’s R&D focus: sensing, perception, control and connectivity that makes such reasoning actionable in the real world with high physical accuracy.

“For the first time, robots can understand complex tasks. ADI delivers the precision physical substrate which, combined with NVIDIA Jetson Thor’s reasoning, responds to real world physics in real time,” says Paul Golding, VP of Edge AI, ADI. “Together, we’re taking humanoids from simulation to shift ready deployment.”

The key to reasoning and physical intelligence

Robotics foundation models compress decades of challenges into perception-rich humanoids capable of dexterous, human-speed manipulation. But their real breakthrough is in reasoning: integrating multimodal inputs to plan, adapt and act in real time.

As noted on our third-quarter 2025 earnings call, ADI’s content opportunity grows with this shift. Every joint needs precise current, position and torque control. Every contact needs tactile and sensory feedback. Humanoids require multiple perception nodes. Each node is a signal chain, perception stack, and power-management opportunity that must run deterministically and with low latency – ADI’s strength.

Closing the Sim2Real gap

ADI is embedding robotics foundation models into the ADI development stack, closing the Sim2Real gap so its hardware behaves in NVIDIA Isaac Sim as it will in the real world. The goal is to build the most physically accurate robotics content in NVIDIA Isaac Sim, enabling teams to iterate at simulation speed and then scale seamlessly to real systems with ADI hardware and NVIDIA Jetson Thor.

Physical intelligence fuses sensing, actuation and policy learning and reasoning so robots can execute precise industrial tasks. It demands high-fidelity edge sensing, energy efficient and functionally safe power, deterministic connectivity to central compute, and a digital twin that closes the Sim2Real loop.

This can now be achieved: NVIDIA Jetson Thor is the compute substrate, and ADI delivers the signal chain fidelity, power integrity, and content that make it actionable.

“With NVIDIA Jetson Thor as the brain and ADI’s high-fidelity sensing, signal-chain fidelity and deterministic connectivity as the nervous system, we take robots from NVIDIA Isaac Sim to the factory floor with physical accuracy – faster,” says Golding

The future of reasoning and physical intelligence

ADI sees growing demand for humanoids across logistics, agriculture and surgical robotics. Frontier use cases include dexterous manipulation of cable assemblies in data centres and automotive manufacturing – tasks that reward speed, precision, and repeatability. ADI’s collaboration on digital twins and policy training in NVIDIA Isaac Sim will address this demand and shorten timelines from concept to production humanoids using ADI’s stack with NVIDIA Jetson Thor.

The same stack – high-fidelity sensing, deterministic connectivity, and digital-twin grounded policy training – extends to other platforms, such as AMRs, where ADI is working with NVIDIA to incorporate ADI perception into cuVSLAM via its IMUs, depth sensors and wheel encoders.