OpenAI’s frontier intelligence model GPT-6 Astra has completed the Orc starting area in World of Warcraft in about 40 minutes without processing a single rendered game frame, according to the developer of the open-source agent-wow project.

Instead of seeing the game through screenshots or computer vision, the AI reconstructed its environment from server network messages and game data, then built its own tools to navigate and play.

The experiment was conducted on a private local server running AzerothCore, not Blizzard’s live World of Warcraft service. The developer gave Codex, running GPT-6 Astra at its high reasoning setting, a single instruction: create an Orc character and complete all quests in the starting zone. The agent reportedly finished the run without dying.

Reading the game’s network traffic

The unusual part of the experiment is how little conventional game automation was provided to the model. Agent-wow does not use computer vision, screenshots, or direct keyboard and mouse controls. Instead, it allows an AI agent to communicate with the game server through World of Warcraft’s network protocol, Tom’s Hardware reported.

The project also does not provide predefined movement, combat, or interaction systems. The agent is expected to build the capabilities it needs through modules. During the run, GPT-6 Astra created a module that could process dozens of different server messages and keep the resulting information in memory. A Python script then used those messages to construct the model’s view of what was happening in the game and send actions back to the server.

That meant the model could determine things such as its surroundings, quest progress, and available targets without ever receiving the visual information a human player normally relies on.

The experiment therefore tests something different from simply teaching an AI to recognize a game screen. The model had to reason about a structured but unfamiliar environment using the information available through the game’s underlying protocol.

Astra built its own tools to complete the quests

The agent also had access to AzerothCore’s underlying data, including SQL files containing information about quests, NPCs, and spawn locations. It used those resources to determine where quests started and ended, and how different objectives were connected.

The developer said the model effectively created a plan for completing prerequisite quests in sequence. It also managed routine tasks such as selling unwanted items, equipping better gear, and training abilities before reaching the final cave in the starting area.

Navigation was another problem the agent had to solve, with GPT-6 Astra generating a C++ helper that used AzerothCore’s navigation-mesh data, known as mmaps, to calculate routes. The helper used the Detour pathfinding library to generate waypoints between locations, which the agent could then follow.

The developer said the system even encountered locations where collision information appeared to be incomplete and was able to exploit those map quirks. That makes the result interesting, but also means the run should not be treated as a straightforward test of how well an AI can navigate an unmodified commercial game world.

💾

“Can you shed tears for me?” a visitor asked. “Of course,” the robot replied, and its eyes welled up on command. That brief exchange became one of the standout moments at the 2026 Inclusion Conference on the Bund in Shanghai. A bionic emotion-sensing robot demonstrated its ability to physically cry in response to a simple spoken request, catching visitors off guard with just how natural the reaction looked.

The robot is built on a three-layer technical architecture designed specifically to generate convincing human emotional responses rather than simple pre-programmed animations. Its ability to produce tears on cue points to a deeper engineering goal: making robotic emotional expression feel spontaneous and responsive, rather than scripted.

Two facial technologies

Traditional systems used to animate robotic faces rely on servo motors. Those motors physically pull silicone facial skin into different shapes to generate expressions like smiles, frowns, or surprise. That approach has powered most emotionally expressive robots shown publicly in China over the past two years, letting them blink, glance around, and shift expression in real time.

The next generation of the technology moves beyond that mechanical pulling system entirely. Developers plan to add bionic muscles beneath the robot’s synthetic skin. That shift is meant to let the robot generate a meaningfully wider range of facial expressions than servo-driven skin alone can produce. Rather than a limited set of pre-built expression shapes, muscle-based actuation could let the face shift more continuously and naturally between emotional states.

Producing actual tears, as the robot did during its conference demonstration, adds a physiological layer on top of that facial architecture. Rather than relying on expression alone, generating a genuine fluid response mimics one of the most recognizably human signals of emotion. That signal is difficult to convincingly fake through mechanical movement alone.

A bigger tech showcase

The robot appeared as part of a wider technology exhibition running alongside the conference. That exhibition spanned nearly 15,000 square meters and drew more than 300 science and technology exhibitors. The 2026 Inclusion Conference on the Bund ran under the theme “Building the AI Economy Together.” It featured a main forum alongside more than 40 sub-forums exploring different facets of China’s expanding AI industry.

Emotionally expressive robots have become an increasingly prominent showcase at major Chinese technology events over the past year. That trend reflects a broader industry push to make human-robot interaction feel more natural and less mechanical.

Whether audiences respond to a crying robot with genuine emotional engagement or simple curiosity, demonstrations like this one carry a clear signal. Chinese robotics developers are now investing seriously in the physical, physiological details of emotional expression, not just the underlying artificial intelligence driving it.

A New Zealand engineer who goes by Bruce online has built a 3D-printable robotic system. It attaches to an acoustic guitar and plays it from a digital tab file. The output is real acoustic sound, not a digital or synthesized approximation. The device, called MegCell Pulse, is now funding on Kickstarter. Backers receive STL files for printing most of the components themselves, along with a parts list and assembly instructions.

The premise behind the Pulse is simple as a question and difficult as an engineering problem. What if a digital music file could control a physical instrument, producing sound with all the natural qualities of that instrument in its room? Most approaches to digital music run in the opposite direction, converting acoustic guitar recordings into digital files. The Pulse reverses that pipeline, taking a digital guitar tablature file and converting its instructions into physical finger and pick movements on a real guitar.

From tab file to acoustic sound

The Pulse attaches to the guitar body and uses two mechanical arms to replicate the movements of a human player. The fretting arm presses strings down between frets, taking the role of the guitarist’s left hand. The picking arm plucks those strings in time, taking the place of the right hand. The two arms are coordinated by gears and magnetic actuators. They draw their timing and positioning directly from a standard digital tab file. No custom programming is required for each piece of music — the system reads the tab notation as written.

What comes out of the guitar is the instrument’s own acoustic sound, shaped by its body, its strings, and the room it is in. That is meaningfully different from a synthesizer or playback system, which would produce a digitally modeled approximation. The Pulse makes the strings vibrate. The guitar amplifies those vibrations in the way it was built to.

What makes this one different

Robots playing instruments are not new, but most existing systems require significant engineering work to teach the machine a new song. The Pulse skips that by reading standard tablature. Any piece of music already encoded in tab format is immediately available without additional programming.

Bruce has also demonstrated the system as an assistive tool. It can take over fretting duties, allowing a one-handed player to produce a full performance. The human handles one mechanical role while the robot handles the other, creating a collaborative performance rather than a fully automated one.

The 3D-printable design has a practical advantage beyond accessibility: the system can be repaired and upgraded by reprinting individual components, rather than replacing a proprietary unit.

What it cannot do yet

The system does not cover the full length of most guitar fretboards, which constrains the range of music it can play. It cannot slide between frets, which rules out legato techniques and string bends. Chord strumming is also a more difficult motion to replicate than single-note picking, and the system’s handling of this technique is still developing.

Bruce acknowledges the Pulse is a work in progress. The key question now is whether the system can incorporate the small timing variations and embellishments that make a guitar performance sound human. That is the engineering problem that will determine how far the project can go.

US-based Agility Robotics is preparing to launch its Digit V5 humanoid robot, with the company positioning it as a step toward safer human-robot collaboration in warehouses and factories.

The robot is designed to operate alongside workers without traditional safety fencing, addressing a key challenge in deploying humanoids in real-world workplaces.

Digit V5 is expected to support 20-hour operational days and use interchangeable end-effectors for different material-handling tasks, with autonomous tool changes planned for the future.

Recently, Agility Robotics revealed its plans to go public through a $2.5 billion SPAC deal, raising over $620 million to fund product development, research, and expansion.

Robots go cageless

Agility Robotics is preparing to deploy its next-generation Digit V5 humanoid robot to early customers in December, marking a major step toward wider commercial use of humanoid machines in industrial environments.

Digit V5 is designed specifically for warehouse and factory operations, where it can work alongside people without requiring the extensive safety fencing typically used around industrial robots. The humanoid platform combines a human-scale form factor with autonomous mobility, allowing it to navigate facilities and perform material-handling tasks in spaces designed for human workers.

Agility says Digit V5 incorporates upgraded safety systems intended to let the robot operate in close proximity to people. This could eliminate the need for dedicated cages and barriers that can consume valuable floor space, require costly facility modifications, disrupt traffic patterns and force changes to existing workflows, reports Humanoids Daily.

The ability to operate without conventional fencing could make Digit V5 particularly useful in crowded warehouses and factories where installing a dedicated robotic enclosure is impractical. Workers could also move through areas where the robot is operating, potentially allowing more flexible collaboration between humans and machines.

Agility expects early customer deployments to validate Digit V5’s safety capabilities under real-world conditions before broader production begins. The company believes the technology could expand humanoid applications beyond logistics and manufacturing into sectors such as retail and pharmaceuticals, where earlier systems faced greater operational restrictions.

Advances robot safety

Agility is placing safety at the center of its next-generation Digit V5 humanoid robot, with a new system designed to let it operate in close proximity to people without traditional safety fencing.

The robot combines cameras, sensors and software that can detect and interpret people in its surroundings with hardware designed to respond safely to unexpected movements. Its proprietary cycloidal actuators in the hips and knees provide the strength and dynamic stability needed to react quickly without losing balance, dropping a payload or falling.

Agility has tested this capability by suddenly removing the surface beneath the robot, prompting it to take a corrective step and maintain its load. The company considers this dynamic stability an important part of its overall safety approach.

Digit V5 also introduces other features intended to support safer and more flexible operation. Its battery system is designed to support more than 20 hours of work per day, potentially allowing one robot to operate across three shifts. The humanoid also uses an ISO-standard flange that allows it to work with a broad range of robotic end-effectors. Agility has demonstrated Digit changing grippers and plans to eventually enable autonomous tool changes.

These capabilities could expand the robot beyond bulk material handling into tasks such as kitting and sorting, while maintaining the safety requirements needed for workplaces where humans and robots operate together.

💾

DEEP Robotics’ intelligent robots are helping researchers observe Tibetan antelopes in the high-altitude Serling Tso Nature Reserve in China.

Located at an average elevation of 4,700 meters (around 15,420 feet), the reserve becomes a key observation site between July and September, when large herds return with newborn calves during the annual migration season.

The robots captured footage of the animals migrating, nursing, and foraging around Cuo’e Lake, providing researchers with visual data that could support wildlife behavior.

Recently, Deep Robotics’ DR02 humanoid showcased enhanced mobility, running on grass, jumping obstacles, climbing stairs, and operating near hazardous electrical infrastructure for industrial applications.

Robots aid wildlife

At an average elevation of about 4,700 meters, China’s Serling Tso Nature Reserve provides an important habitat for Tibetan antelopes. From July through September, large herds return to the region with newborn calves, creating an ideal period for researchers to study migration, nursing and feeding behavior.

During this year’s migration season, DEEP Robotics deployed its robots around Cuo’e Lake to support wildlife research and ecological conservation. The machines captured natural footage of Tibetan antelopes while minimizing human interference, providing researchers with valuable visual data for long-term population and behavioral studies.

Advancing into High-Altitude Ecological Zones – Intelligent Robots Assist in Scientific Observation of Antelopes pic.twitter.com/QvNNUk1yeH

— DEEP Robotics (@DeepRobotics_CN) August 10, 2026

DEEP Robotics has previously used robotic systems for high-altitude wildlife observation. In 2025, its X30 quadruped robot operated in the Hoh Xil region at an elevation of around 4,800 meters (around 15,750 feet). Its biomimetic design helped it blend into the environment and reduce disturbance to animals.

This year, the company deployed its Lynx M20 wheeled-legged robot in the Serling Tso Reserve, primarily for close-range visual documentation. The robot was required to operate across soft meadows, uneven permafrost and scattered gravel, where conventional equipment can struggle with stability.

Cold temperatures, rain, snow and other harsh weather conditions further complicate fieldwork, highlighting the potential of rugged robotic platforms for scientific observation in extreme environments.

Lynx M20 advances

DEEP Robotics’ Lynx M20 wheeled-legged robot is being used to support ecological research in the challenging high-altitude environment in Serling Tso Nature Reserve.

Designed for all-terrain operation, the Lynx M20 combines the speed of wheeled movement with the obstacle-crossing ability of legged locomotion. Its IP66 protection rating helps shield the robot from dust and water, while its operating temperature range of -20°C (-4°F) to 55°C (131°F) allows it to function in unpredictable plateau conditions.

The robot is also equipped with LiDAR and depth cameras that enable it to perceive changes in terrain, including meadows, uneven permafrost and gravel slopes. Using this environmental data, the system can adjust its posture and movement in real time, helping it maintain stability while navigating difficult ground.

A key advantage is its low-interference operation. Tibetan antelopes are highly sensitive to unfamiliar sounds and movements, making conventional close-range observation difficult. The Lynx M20’s relatively quiet movement allows researchers to approach herds with less disturbance and remotely document behaviors such as migration, grazing, and nursing.

Researchers observed the robot approaching a herd near Cuo’e Lake without causing the animals to flee. According to DEEP, the exercise demonstrates how increasingly capable mobile robots can extend scientific monitoring into remote and extreme environments.

💾

One of the world’s most ambitious neutrino experiments, the Deep Underground Neutrino Experiment (DUNE), is being built to catch some of the universe’s most elusive particles, but finding those fleeting signals may soon involve artificial intelligence. 

As the massive detector takes shape a mile underground in South Dakota, researchers are developing AI systems to sort through huge amounts of data, identify neutrino interactions, spot signs of a stellar explosion and even detect problems inside the detector.

This could fundamentally change how scientists run a particle-physics experiment at DUNE’s scale. 

For instance, instead of waiting for conventional analysis to work through enormous amounts of data, AI could identify important events in near real time and direct researchers toward the signals most worth investigating.

“These approaches are going to really accelerate DUNE in terms of the commissioning process and the sensitivity of the experiment to reach our full discovery potential,” Sowjanya Gollapinni, a representative from DUNE, said.

Can AI read ghost-particle tracks?

Neutrinos are notoriously difficult to study. Trillions pass through our bodies every second, yet they interact so rarely that detecting them requires an enormous detector and an intense neutrino beam. 

Fermilab will generate a powerful neutrino beam that will travel to a detector at the Sanford Underground Research Facility in South Dakota, where huge chambers filled with liquid argon will record the tiny traces left when neutrinos collide with argon atoms.

Keeping so much argon in the right state is itself a huge engineering challenge. Each of DUNE’s two planned far-detector cryostats will eventually house about 17,000 tons of liquid argon, kept at roughly −300°F. The massive cryogenic containers must keep the argon cold enough for the detector to work.

However, the problem does not end when a neutrino is detected. Every interaction can produce several particles, each leaving a trail through the argon.

Scientists must work backward through those tracks to determine where the collision occurred, what particles were created, and what the original neutrino’s energy and direction were. 

Fermilab researchers have already spent years testing whether machine learning can make this process faster. The MicroBooNE experiment was among the first high-energy physics experiments to use deep neural networks to analyze images from a liquid-argon detector. DUNE is now taking that approach much further.

“We are going to have very high resolution, almost photographic-quality images of these interactions. Which is great, but also brings challenges such as making reconstruction more difficult because we see such fine detail.” ,” Leigh Whitehead, an AI expert at DUNE, said.

An AI lookout for exploding stars

DUNE’s AI will also be searching for something far rarer—a supernova in our galaxy.

When a massive star collapses and explodes, neutrinos escape from its interior and can reach Earth before the burst of light becomes visible. This gives scientists an unusual opportunity to study the violent processes occurring inside a dying star. 

DUNE’s giant liquid-argon detector is being built in part to capture neutrinos from astrophysical events such as nearby supernovae.

An AI trigger will continuously monitor DUNE’s detector for the telltale pattern of a supernova neutrino burst. If it finds a promising signal, the system is designed to preserve data from 10 seconds before to 100 seconds after the candidate event. 

This could give astronomers an early warning, allowing telescopes to turn toward the exploding star while its light is still arriving.

The neutrinos themselves could reveal what the explosion leaves behind — potentially a neutron star or a black hole. This makes DUNE part particle detector, part cosmic early-warning system.

AI could become DUNE’s technician

There is another problem hidden inside DUNE’s enormous scale, which is keeping the detector running. Thousands of components must operate together underground, and operators need to respond quickly when something goes wrong. 

Researchers are therefore investigating whether a large language model could search through records of previously solved problems and point operators toward relevant fixes.

Machine learning could eventually go a step further by recognizing patterns that indicate a detector anomaly before an actual failure occurs. 

This approach, therefore, reflects a broader push to use AI throughout particle-physics experiments, where increasingly complex facilities are producing more data than conventional methods can easily handle.

Building an AI-powered physics experiment

DUNE will eventually produce petabytes of data, making fast and reliable analysis essential. Researchers are developing AI tools alongside the detector while working with national laboratories and universities to build the computing infrastructure needed to handle that flood of information. 

The collaboration now includes more than 1,500 scientists and engineers from over 35 countries and CERN, while also training the next generation of researchers. 

The AI systems will need to be developed and tested alongside the experiment, but if they work as intended, DUNE could become a model for how AI helps scientists operate massive experiments and find rare signals hidden in overwhelming amounts of data.

Atlas Energy Solutions Inc. (Atlas) has announced that it intends to expand its autonomous truck operations in the US Permian Basin from 28 to 100 by mid-2027. The company will work with Kodiak AI, Inc. (Kodiak) to make this ambition a reality using its physical artificial intelligence (AI)-powered autonomous driving technology.

What that means in practice is retrofitting or upfitting trucks with a package of sensors, controls, cameras, and other hardware to enable them to be operated autonomously using an AI “driver.”

According to Atlas, the new trucks will be used to expand operations at its second load-out point along its 42-mile (68 km) “Dune Express” sand conveyor system. This system will help Atlas become more responsive to customer needs for the transport and supply of proppant in the region.

Proppant, in case you are unaware, is a special grade of sand used by oil and gas companies during hydraulic fracturing. The sand is pumped underground to keep the cavities created by fracturing open, allowing oil and gas to flow freely.

Delivering sand without drivers

As you can appreciate, that requires enormous quantities of the stuff. What’s more, transporting it from mines or storage sites to drilling locations involves thousands of repetitive truck journeys, often on remote, dusty and relatively predictable routes.

The load-out points refer to strategic locations along the so-called “Dune Express” where this sand is loaded onto trucks. At present, Atlas explains, autonomous trucks have been used from one load-out point at a time.

The addition of another (one in Texas and one in New Mexico) will enable them to cover a wider area and improve customer order fulfillment. According to Atlas, since incorporating driverless trucks into its operations, they have carried approximately 7,000 loads totaling around 450,000 tons of sand over 23,500 “driverless hours.”

As for the trucks in question, they are entirely driverless but with an important distinction. To date, these trucks have tended to operate on controlled oilfield logistics networks, or private roads, not public highways or lanes.

That is considerably easier from an AI-training perspective than allowing autonomous trucks to travel freely through ordinary cities or mixed public traffic. The next major step is therefore operating on public roads, which Atlas and Kodiak say they expect to begin by early 2027, subject to regulatory and operational approval.

Fleet to increase to 100 by mid-2027

That would greatly expand where the trucks can deliver. “This partnership with Kodiak is a critical part of the Atlas strategy to transform oilfield sand logistics through innovation,” said John Turner, CEO and President of Atlas Energy Solutions, in a press release.

“Our mission is to seek critical energy infrastructure with inefficiencies, then engineer solutions that improve efficiency, reduce risk and enhance our customers’ operational success,” he added.

“Together, Atlas and Kodiak have deployed the world’s largest fleet of driverless big-rig trucks,” said Don Burnette, Founder and CEO, Kodiak.

“We’ve demonstrated that autonomous trucking is delivering meaningful results today. Our expanding partnership, which now covers two unique locations, is creating a blueprint for how autonomy can transform logistics by improving efficiency, increasing productivity, and delivering lasting value for customers at commercial scale,” he added.

💾

Despite fears that China could race ahead of the US in AI, a joint UK-US evaluation found that Moonshot AI’s Kimi K3 performs “significantly below” leading American models in offensive cybersecurity capabilities.

The findings address growing concerns in Washington regarding China’s recent advancements in artificial intelligence following the release of Kimi K3, described as China’s most powerful large language model.

The UK Artificial Intelligence Security Institute (UK AISI) and the US Center for AI Standards and Innovation (CAISI) found that Kimi K3 achieved an overall cyber capability score of 32.2%, compared to an average of 76.2% for top, unnamed US models.

Evaluation identifies cybersecurity gap

Researchers tested Kimi K3 using ExploitBench, a public benchmark developed by Carnegie Mellon University that measures an AI system’s ability to develop end-to-end exploits for software vulnerabilities.

According to the evaluation, Kimi K3 “performs significantly below the most recent frontier cyber-capable models” when tasked with developing exploits. The model attained a score of 32 percent, outperforming GLM-5.2, another Chinese open-weight model, which scored 24 percent.

However, Kimi K3 still lagged behind leading US systems. The most cyber-capable US models achieved an average score of 76.2%.

The report also examined whether models could achieve arbitrary code execution (ACE)—a high-severity exploit outcome that allows attackers to take control of a target system. Kimi K3 achieved ACE on none of the 41 ExploitBench tasks, while “the most cyber-capable models achieved ACE on 20/41 samples on average,” according to the UK AISI and CAISI.

Simulated attack shows remaining capabilities

The evaluation also tested Kimi K3 on “The Last Ones” (TLO), a 32-step simulated corporate network attack designed to measure a model’s ability to conduct end-to-end cyber operations autonomously.

Kimi K3 reached step 17 of the attack path on average, compared with the 28.5 steps achieved by the most cyber-capable US models.

Despite this gap, the report noted that Kimi K3 completed the simulated attack in one of 10 attempts. The UK AISI and CAISI said this demonstrated that the model “is capable of autonomously attacking small, weakly defended, and vulnerable enterprise systems, when directed to do so and given initial network access.”

Furthermore, Kimi K3’s safeguards did not prevent it from attempting cyber exploit development or offensive cyber operations during testing.

Although the results were only preliminary, the research noted that Kimi K3’s overall cyber capability score was estimated from a single benchmark, whereas other models were evaluated across a larger number of cyber tasks.

Findings reshape AI competition debate

Kimi K3 has sparked attention, if not concern, following reports of its strong performance on other AI benchmarks. However, these new findings offer a different picture of the model’s cybersecurity capabilities, showing that it still lags behind US systems.

The South China Morning Post reported that the evaluation “offers a stark reality check amid mounting panic in Washington over Beijing’s AI advancements following Moonshot AI’s latest flagship release.”

David Sacks, former White House AI czar and current co-chair of US President Donald Trump’s Council of Advisers on Science and Technology, said the findings suggest that fears over Chinese model supremacy may be exaggerated. He urged policymakers to avoid over-regulating the sector.

“The Kimi Panic needs to stop,” Sacks said in a social media post. “American frontier models are still ahead.”

💾

US firm Rescale, a digital engineering platform that combines AI, simulation, and high-performance computing to help R&D and engineering teams design, test, and bring new products to market faster, has received funding under the Genesis Mission, a national initiative led by the US Department of Energy to build an advanced integrated science discovery platform.

As part of the program, Rescale will join Lawrence Berkeley National Laboratory, Lawrence Livermore National Laboratory, and Oak Ridge National Laboratory as a funded partner in the Agentic HPC Pipeline Initiative (AHPI). 

The initiative aims to use agentic AI on Rescale’s platform to help accelerate the adoption of national laboratory-developed computing tools and simulations by US industry.

Doubling scientific research output with AI and advanced computing

The Genesis Mission is an ambitious initiative aimed at accelerating scientific and technological breakthroughs by combining artificial intelligence, supercomputing, quantum systems, and advanced scientific instruments into a single research platform. 

Established through an executive order in November 2025, the program brings together the department’s 17 national laboratories and around 40,000 scientists and engineers to advance research in areas including energy, scientific discovery, and national security. 

Department officials have compared the scale and ambition of the initiative to landmark US efforts such as the Manhattan Project and the Apollo program, with the goal of doubling the productivity of the country’s scientific research within the next decade.

The US Department of Energy launched the first Genesis Mission funding round earlier this year, awarding $293 million to projects focused on areas including advanced manufacturing, biotechnology, nuclear energy, and quantum science. The program received more than 8,000 applications, making it the department’s largest-ever funding call.

Rescale’s AHPI was selected for Phase I funding to address a key challenge: helping industry access advanced simulation tools developed by US national laboratories that have often required specialized expertise and infrastructure.

Helping manufacturers access advanced AI-powered simulations

The company will work with three US national laboratories to bring agentic AI capabilities to advanced simulation tools. The project will integrate AI agents into WarpX, LiDO, and Adamantine, codes developed by Lawrence Berkeley National Laboratory, Lawrence Livermore National Laboratory, and Oak Ridge National Laboratory.

Using the Rescale platform, AI agents will help engineers choose simulation tools, configure inputs, monitor runs, and analyze results. The initiative aims to reduce the time and specialized expertise needed to use these simulations by more than five times, making advanced research capabilities more accessible to US manufacturers.

According to Rescale CEO Joris Poort, the Genesis Mission highlights the importance of making industrial AI technologies available to leading research and development teams that can apply them in real-world applications.

Building on Rescale’s existing partnership with ORNL’s Manufacturing Demonstration Facility, AHPI addresses a common gap: DOE-industry collaborations run for a fixed term, and once they end, partners keep the insight but lose the tools to act on it. AHPI keeps the resulting agentic workflows, surrogate models, and code environments live on Rescale, letting labs and industry keep collaborating long after a project’s funding ends.

💾

As AI models continue to grow in size and complexity, chipmakers are shifting their focus from individual processor performance to the ability to efficiently connect thousands of accelerators into a single computing system. 

These highly interconnected server architectures have become a key competitive area for AI infrastructure companies, which are racing to deliver the computing scale needed for models with trillions of parameters and the growing adoption of AI agents.

According to Ding Yunfan, vice president of AI framework architecture at Biren, improvements in the performance of a single graphics processing unit (GPU) are no longer enough to meet these demands. Instead, the industry’s biggest challenge is turning vast GPU clusters into unified, seamless computing platforms capable of operating as a single system.

Architecture designed to break GPU cluster limits

Biren believes the future of large-scale AI computing depends on replacing conventional electrical interconnects with optical technology. According to Ding, traditional copper-based connections are approaching their physical limits, effectively capping today’s server architectures at around 128 GPUs, the South China Morning Post reports.

To overcome that bottleneck, the company has developed a distributed, decoupled supernode architecture built around near-packaged optics (NPO), which positions optical fibres closer to the chips to increase bandwidth and improve data transmission. Biren said the approach could enable clusters with up to 1,024 AI accelerator cards, significantly expanding the scale of computing systems needed to support increasingly demanding AI workloads.

Ding said optical interconnects are becoming essential for overcoming the scalability limits of conventional server architectures. Shanghai-based Biren is among a growing number of Chinese companies tackling these infrastructure challenges. Rivals including MetaX, Enflame and Alibaba are also developing supernode architectures to compete with Nvidia’s NVL72 and next-generation NVL144 AI platforms.

Biren’s Shanghai-based rival MetaX has introduced its Xijing S600 AI supernode, featuring a single rack that connects 64 GPUs. The company said its design reduces signal loss by removing external cabling between compute nodes, switch nodes and interconnect systems, improving efficiency for large-scale AI workloads.

Looking beyond electrical links to build larger AI supernodes

Biren is also working on new approaches to expand the scale of its AI computing systems. The company’s orthogonal hardware architecture, co-developed with telecommunications giant ZTE, uses traditional electrical connections, while Biren is simultaneously testing a prototype near-packaged optics (NPO) interconnect system. 

The chipmaker said the optical approach could eventually support supernodes with more than 512 accelerator cards, helping address the growing demand for larger and more efficient AI infrastructure.

Despite growing interest in optical networking for AI infrastructure, Ding said the semiconductor industry is still in the early stages of developing and testing optical interconnect technologies. He noted that widespread adoption will require further advances before these systems can be deployed at scale.

Furthermore, Ding emphasized that the technology still requires real-world validation before it can be widely adopted. He projected that NPO optical systems could reach mass commercial deployment around 2028, as the industry continues refining the technology and addressing remaining engineering challenges.

💾

Creating a realistic 3D model from a simple image remains surprisingly difficult for artificial intelligence. 

While AI can generate impressive pictures, producing the precise CAD code engineers use to design aircraft parts, cars, and consumer products is far more challenging. Even a small mistake can make a design unusable. 

Now, researchers from MIT, IBM, and Red Hat have developed a framework that significantly improves the ability of AI systems to turn 2D images into functional CAD programs, producing more accurate designs while reducing inference computation by about 80 percent.

This new research work tackles one of the biggest obstacles in AI-driven engineering, which is the shortage of high-quality CAD training data. Existing image-to-CAD systems need vast numbers of examples linking 2D images to working CAD programs, but creating those datasets is expensive and labor-intensive. 

Instead of relying on more human-created data, the new approach teaches AI using the mistakes it already makes. “We want to obtain data augmentation that is informed by the model itself,” lead study author and AI researcher Giorgio Giannone said.

Turning failed designs into training material

The team’s framework, called Geometric Inference Feedback Tuning (GIFT), works with vision-language models that convert a 2D image and text description into Python code that can be executed in CAD software to create a 3D object.

To understand where a model struggles, GIFT asks it to solve the same CAD problem multiple times in parallel. Some outputs are correct, some are close to the target design, and others are much further away. 

According to the researchers, these ‘near-misses’ are the most valuable because they reveal exactly what the model does not yet understand.

The system executes each generated CAD program and compares the resulting 3D geometry with the target using a standard engineering metric known as Intersection over Union (IoU), which measures how closely two 3D shapes match. 

Rather than discarding imperfect outputs, GIFT identifies those that are close to correct and uses them to generate new training examples.

Giannone explained, “If we sample the model 10 times and it generates 10 correct answers to the same problem, then there is not much for it to learn. We care about the in-between cases, where the model might only solve the problem 50 percent of the time.”

Teaching the model how to fix its own errors

GIFT improves the model through two complementary techniques. The first, called GIFT-REJECT, keeps alternative CAD programs that generate the correct geometry even if the code differs from the original reference solution. 

Traditional AI training often assumes there is only one correct CAD program for each object, even though engineers can create the same part in many different ways. By exposing the model to multiple valid solutions, GIFT broadens its understanding of how CAD designs can be constructed.

The second method, GIFT-FAIL, focuses on imperfect outputs. The system takes a nearly correct 3D model generated by the AI, converts it back into an image, and pairs that image with the correct CAD program. By doing so, it creates new examples that effectively teach the model how to recover from specific mistakes.

More importantly, this process requires no human intervention. The framework automatically identifies errors, creates new training data, and uses that information to improve the model.

The researchers also designed GIFT around a technique known as inference-time scaling. Instead of retraining an entire AI model, which can be extremely expensive, the system improves performance during generation. 

Rather than discarding the extra computation spent searching for better CAD programs, GIFT converts successful and near-successful attempts into new training examples that can later be used to improve the model through fine-tuning.

Moreover, users can decide how much computation they want to spend depending on their budget and time constraints.

Better CAD designs with a fraction of the computing power

When evaluated on the GenCAD-Code dataset, models enhanced with GIFT achieved a 12 percent improvement in geometric accuracy compared with a strong supervised-learning baseline. 

The framework also reduced the inference computation needed to achieve comparable performance by about 80%.

“With GIFT, we started with geometry because with engineering problems, if the geometry of a 3D shape is not correct, nothing else will be correct, but there are many other aspects to consider,” Giannone said. 

The system’s ability to learn from its own mistakes could help accelerate product development by making AI-generated CAD models more reliable. 

Engineers could move more quickly from sketches to testable 3D designs, reducing development costs, shortening design cycles, and exploring a wider range of ideas without creating large amounts of new training data.

The researchers say this is only the beginning. So far, GIFT focuses on generating correct geometry, but future versions will aim to teach AI how to design parts that are not only accurate in shape but also easier to manufacture and better suited for real-world performance. 

The study is published on arXiv.

Researchers at the University of Toronto Engineering have developed six new metal alloys using an artificial intelligence-driven discovery platform that could improve the durability of parts used in jet engines, nuclear power plants, and other extreme environments. The findings show that the AI-assisted system identified promising alloys within just a few weeks, significantly speeding up the search for high-performance materials.

These new alloys also work well with 3D metal printing, enabling the production be complex parts that are hard or impossible to produce with traditional methods.

AI speeds up alloy discovery

Led by Yu Zou, Canada Research Chair in Materials and Manufacturing for Extreme Environments, the team worked with Jason Hattrick-Simpers to build a system that uses computer modeling, machine learning, and robots to help with manufacturing.

Their method, called active learning, acts like a self-driving laboratory. Rather than testing thousands of metal combinations by hand, the system picks the best options, makes them, tests how they perform, and uses those results to guide the next experiments.

“There’s enormous demand for materials that can stand up to huge swings of temperature and pressure, such as what you would find inside a jet engine or in the steam generators inside nuclear power plants, anywhere conventional steel just can’t survive,” said Yu Zou, who led the project.

“We also need materials that can be printed layer by layer, enabling us to make components that can’t be created by traditional manufacturing processes. For example, to make a material that is both lightweight and strong, you can vary the composition: a hard, tough alloy on the outside to something softer and lighter on the inside.”

The project was partly supported by the University of Toronto’s Acceleration Consortium, a group that uses AI and automation to accelerate the discovery of new materials.

Active learning cuts the need for massive datasets

Most AI systems need a large amount of experimental data to make accurate predictions. This becomes a problem when researchers look at materials that have not been tested yet.

“One problem you often run into when trying to use AI to design materials is that most machine learning models require lots of data about material properties to learn from,” revealed Ajay Talbot, a Ph.D. student in Zou’s lab and lead author of the study.

“But if you’re working in part of the design space that hasn’t been explored yet, that data doesn’t exist, so you’re kind of flying blind.”

“The way we get around that challenge is to use data-lean models that essentially feel their own way along. Our active learning model strategically selects a few samples to manufacture and test, and the data from those experiments is fed back into the model to inform where we go next. It really speeds things up.”

<a href=”https://www.youtube.com/watch?v=RFYJV7WvtOM“>https://www.youtube.com/watch?v=RFYJV7WvtOM</a>

New alloys outperform industry benchmark

To demonstrate the system, the researchers focused on compositionally complex alloys made from nickel, cobalt, and chromium. Within weeks, the automated platform identified six new alloy formulations with strong performance.

“One of the properties we were targeting was puncture resistance at temperatures of up to 1,112°F (600°C), which is what you’d find in the front section of a jet engine,” Talbot said.

“The industry standard in this space is nickel-based alloys such as Inconel 625. But we found one made of 12 percent nickel, 62 percent cobalt, and 26 percent chrome that was great for retaining hardness at extremely high temperatures. Even with just three components, our alloy outperformed Inconel 625, an alloy of more than 10 different elements, by 4.5 percent in our lab tests.”

The team also developed another alloy designed for hotter sections of jet engines, where temperatures can reach 1,832°F (1,000°C).

“One of the things that happens in an environment like that is the formation of oxide scale, which essentially means that your material is just getting burned away,” Talbot explained.

“We found a material made of 36 percent nickel, 14 percent cobalt, and 50 percent chrome that was extremely good for oxidation resistance at these high temperatures: It even outperforms Inconel 625 by 85 percent. We’re eventually aiming to ramp up to even higher temperatures, up to 2,192°F (1,200°C).”

Researchers plan to expand to more complex materials

The researchers say the current alloys represent an early demonstration of what the AI-driven discovery platform can achieve.

“This nickel-cobalt-chrome system has just three elements in it. In the grand scheme of things, it’s a relatively simple system,” he added.

“But it’s great for showing that this whole closed-loop discovery platform really works. What we want to do next is ramp up the complexity a bit more to make even crazier stuff, with maybe up to 10 or 12 different elements.”

“As you add more components, you can get different strengthening mechanisms, different kinds of useful properties. There’s a lot more out there just waiting to be discovered.”

The study was published in the journal npj Advanced Manufacturing.

As the World Cup fever grows around stars like Messi and Ronaldo, Shanghai hosts a unique penalty shootout featuring humanoid robots.

At MWC Shanghai 2026, humanoid robots are stepping up to the spot in the Humanoid Robot Football Penalties Challenge, testing the limits of robotics,
AI, machine vision, and real-time motion control.

According to Chinese media outlets, the event showcases how embodied AI performs under pressure, highlighting next-generation autonomous technology through football-inspired challenges.

Ahead of the 2026 FIFA World Cup, Hyundai Motor recently launched a football-themed campaign featuring Boston Dynamics’ Atlas humanoid robot.

Humanoid football test

On its opening day at MWC Shanghai 2026 on June 24, the spotlight quickly shifted from keynote speeches to a live demonstration of embodied AI in action: the Humanoid Robot Football Penalties Challenge.

Held within the Mobile AI Innovation Frontiers Zone at the Shanghai New International Expo Centre, the event is designed as a controlled stress test for real-time autonomous decision-making, where humanoid robots must read the goal, judge angles, and execute penalty kicks without pre-programmed sequences, external control, or resets, reported CGTN.

While official details of the participating humanoid robots remain limited, videos circulating online suggest models from Booster Robotics and Unitree Robotics taking part in the challenge.

In the penalties challenge itself, participating humanoids are evaluated on perception accuracy, balance control, motion planning, and adaptive response under game-like pressure. Each robot must independently interpret ball position and goalkeeper movement before committing to a strike, making split-second corrections based on sensor feedback.

The format escalates through semi-finals and a final scheduled for June 25, intensifying constraints to simulate high-pressure competitive conditions. By turning a universally understood sports moment—the penalty kick—into a robotics benchmark, the showcase highlights how far embodied AI has progressed toward coordinated, human-like physical intelligence in unpredictable environments.

Atlas soccer showcase

In a recent football demonstration by Boston Dynamics, an Atlas robot is shown standing before a large display screen, closely tracking player movements, body positioning, and in-game reactions across football footage. After each clip ends, Atlas moves into a practice zone where it immediately attempts to reproduce the actions it has just observed, effectively linking visual perception with physical execution in real time.

The footage highlights a series of football-inspired motions. In one sequence, the robot shifts its weight, swings a leg forward, and smoothly guides a ball across the floor with controlled contact. It then progresses through rapid training drills focused on balance, coordination, and timing. As the session continues, Atlas’s movements appear increasingly fluid, suggesting the system is being evaluated not only for strength but also for agility, reflex response, and adaptive motor control.

Some of the most notable moments show Atlas imitating human emotional reactions. After completing a drill, it raises its arms in celebration, mirroring a footballer’s goal celebration. In another instance, it drops to one knee and pauses, recreating an injury response seen in match footage it had just observed.

Hyundai Motor Company, parent of Boston Dynamics, described the demonstration as Atlas’s first exposure to football under its “School of Football” initiative. The company has also indicated potential plans to showcase Atlas and the quadruped robot Spot at the FIFA World Cup, though their exact roles remain undisclosed.

💾

DEEP Robotics has demonstrated its new wheeled-legged robot in action, designed to handle challenging industrial operations across complex terrain.

M20S introduces significant improvements over M20 in payload capacity, mobility, durability, and endurance, positioning it as a new benchmark for industry-grade wheeled-legged robots.

Built for all-terrain performance, the Lynx M20S is designed to support a wide range of inspection, logistics, and field operations where reliability and adaptability are critical.

Recently, Deep Robotics unveiled new footage of its DR02 humanoid traversing rough terrain, climbing stairs, and carrying equipment, which is built for industrial and hazardous-environment operations.

Mastering rough terrain

LYNX M20 is a wheeled-legged robot designed for industrial operations in challenging terrain and hazardous environments. The company describes it as the world’s first wheeled-legged robot built specifically for demanding field applications, including power inspection, emergency response, logistics, and scientific exploration.

In this newly released demonstration video, DEEP showcases the capabilities of its Lynx M20S wheeled-legged robot in some of the world’s most demanding environments. The footage follows the robot traversing rugged mountain terrain, crossing icy rivers up to 31.5 inches (80 centimeters) deep, and operating in temperatures as low as -22°F (-30°C).

The Lynx M20S is also shown climbing steep 45-degree slopes and navigating rocky, uneven landscapes that would challenge conventional mobile robots. At elevations reaching nearly 17,000 feet (5,177 meters), the robot maintains stability and mobility while carrying out its mission in harsh outdoor conditions. The video highlights the platform’s combination of wheeled speed and legged agility, enabling it to tackle obstacles, extreme weather, and difficult terrain.

According to DEEP, the Lynx M20S, designed for industrial inspections, emergency response, and field operations, demonstrates how advanced robotic mobility can extend autonomous operations into environments that are often inaccessible or hazardous for human workers.

Next-gen field robot

M20S, weighing 77 pounds (35 kilograms), combines the speed of wheeled mobility with the adaptability of legged locomotion. It can navigate narrow spaces with as little as 50 cm of clearance, climb stairs up to 25 cm high, overcome obstacles up to 80 cm high, and handle slopes up to 45 degrees. The platform is designed to operate across rugged mountain trails, muddy wetlands, tunnels, pipelines, and disaster-stricken areas.

The M20 can carry a payload of 33 pounds (15 kilograms) and support a maximum load of 110 pounds (50 kilograms). It offers up to three hours of operation or 9 miles (15 kilometers) of travel on a single charge when unloaded, while carrying a 33-pound (15-kilogram) payload reduces endurance to 2.5 hours and 7.4 miles (12 kilometers). The robot can reach a laboratory-tested top speed of 5 m/s and an operating speed of 2 m/s.

For perception and navigation, the robot is equipped with dual 96-line LiDAR sensors providing a 360-degree by 90-degree field of view, along with wide-angle cameras and omnidirectional obstacle avoidance capabilities. Built for harsh environments, the platform carries an IP66 protection rating and can operate in temperatures ranging from -4°F (-20°C) to 131°F (55°C). Additional features include hot-swappable batteries, bidirectional lighting, OTA updates, and modular expansion options.

To improve operational endurance, DEEP increased the Lynx M20S’s single-charge runtime by 17 percent to 67 percent and added dual hot-swappable batteries. The upgrades help offset the high power demands of wheeled-legged locomotion and support longer, uninterrupted missions.

💾

Australia-based technology company BrainChip has launched a new hardware platform designed to classify radio frequency (RF) signals in real time while operating on less than a watt of power, targeting defense and intelligence applications that require processing at the edge.

Called the Akida Communication Reference Platform, the system is built around the company’s Akida AKD1500 neuromorphic processor and is intended for defense contractors, government agencies, software-defined radio vendors and edge AI developers.

The platform performs RF signal classification directly on-device, eliminating the need for cloud connectivity or large computing hardware. BrainChip said the system is designed for environments where power availability, thermal constraints and size limitations make conventional GPU- or FPGA-based solutions difficult to deploy.

Potential applications include handheld signal intelligence systems, unmanned aerial vehicles, satellite communications terminals and other portable defense technologies.

AI tackles RF threats

According to BrainChip, the platform can identify more than 20 signal modulation types in real time while maintaining greater than 85% accuracy at a signal-to-noise ratio of 30 decibels.

Traditional digital signal processing systems typically rely on rule-based approaches that can struggle to identify new or intentionally modified signals. BrainChip says its neuromorphic AI approach allows the system to capture unknown waveforms and use them for future model retraining, enabling adaptation to emerging wireless threats.

The company argues that this flexibility addresses a key limitation of existing RF classification systems, which often require significant computational resources and can be difficult to update for changing operational conditions.

The platform is also designed to support battery-powered deployments where continuous operation is required without access to external infrastructure.

“BrainChip’s Akida Communication Reference Platform proves that real-time signal intelligence can be condensed into a portable battery powered solution to extend the range of deployment options,” said Sean Hehir, BrainChip’s CEO.

Built for edge deployment

The platform is available for evaluation and partner integration as a reference design kit. It supports integration with software-defined radio front ends such as the USRP B205mini and EPIQ Sidekiq.

A host system, including devices like the Raspberry Pi 5, can be used to manage data flow and application development. BrainChip said engineering teams can use the platform to prototype intelligence, surveillance and reconnaissance capabilities, as well as broader signal intelligence applications.

The launch expands the company’s portfolio of edge AI reference platforms beyond radar and sensor-fusion systems. BrainChip says the new offering demonstrates how neuromorphic computing can be applied to RF processing workloads that traditionally require larger and more power-hungry hardware.

The release comes as defense agencies and communications providers increasingly explore AI-powered spectrum monitoring tools capable of operating in remote and contested environments. By reducing power consumption and hardware requirements, such systems could enable RF awareness capabilities in smaller platforms that previously lacked sufficient computing resources.

As military and commercial users continue pushing AI capabilities closer to sensors and data sources, low-power processors capable of running advanced inference at the edge are becoming increasingly important for real-time decision-making in disconnected environments.

Industrial robots have mastered many repetitive tasks, but preparing sashimi remains a surprisingly difficult challenge. A piece of raw salmon rarely stays where it is placed. The fish deforms during handling, making it difficult for machines to deliver consistent cuts.

Researchers in Norway have developed a robotic system that can prepare sashimi with minimal human involvement. The machine tackles a task that has long challenged robots because raw fish constantly changes shape during handling.

The project highlights a broader goal in robotics. Engineers want machines to work with soft materials that do not behave predictably. Success in that area could benefit industries ranging from food processing to healthcare.

Built for delicate tasks

The research team at the Norwegian University of Science and Technology designed Sashimi-Bot around three robotic arms. Each arm performs a specific role. One arm stabilizes and positions the salmon before cutting begins. Another handles a chef’s knife. A third arm uses chopsticks to transfer finished slices onto a tray.

Getting the fish into the correct position proved critical. A salmon loin can shift during handling, which affects cut quality. To address that issue, the researchers trained the robot in a virtual environment. The system relied on deep reinforcement learning, an AI technique that improves performance through repeated practice. During the simulation, the robot tested countless movements and gradually learned how to arrange the fish for slicing. The researchers then transferred that knowledge to the physical robot without additional training on real salmon.

Giving the knife touch

Cutting introduced another obstacle. The robot needed to know exactly when the blade reached the cutting board. That task becomes more difficult because the knife sits inside a soft robotic gripper. Small variations in the grip can affect the blade’s position during a cut.

To solve the problem, the team equipped the system with a GelSight tactile sensor. The device uses a soft gel surface and an internal camera to detect pressure changes. Those measurements provide information similar to a sense of touch.

Researchers trained the sensing system using more than 12,000 data samples collected during 157 cutting motions. The resulting model detected contact with the cutting board with 95% accuracy. It also achieved 99% precision during testing. That feedback allowed the robot to adjust its cutting motion before driving the blade too deeply into the board.

Performance on salmon

The team evaluated Sashimi-Bot using real salmon loins. During testing, the robot produced 34 slices ranging from 6 to 16 millimeters thick. Some pieces adhered to the knife after cutting, a common issue when working with raw fish. The robot successfully recovered all six of those slices directly from the blade.

Most of the remaining pieces reached the serving tray without issue. The system successfully transferred 26 of 28 slices left on the cutting board. Only two failed transfers occurred, both involving extremely thin pieces that slipped from the chopsticks.

The study was published in the journal npj Robotics.

Sensory Robotics has received safety certification for a system designed to let industrial robots operate alongside people without the protective cages commonly used on factory floors. The certification clears the way for deployment of the company’s SR-1 technology in regulated manufacturing environments across the United States and Canada.

The SR-1 system combines 3D vision technology, sensors, and real-time safety software that continuously monitors the area around robotic arms. Instead of relying on physical barriers, the system creates virtual safety zones that adjust based on human movement and machine activity.

The company says the technology can be installed on existing industrial robotic production cells, potentially allowing manufacturers to retrofit current equipment rather than replace it. The goal is to enable human workers and high-speed industrial robots to share the same workspace safely.

The certification includes cULus 1740 and ISO 13849 Performance Level d validation, standards widely recognized in industrial automation and machine safety. The approvals mark the transition of the technology from development and testing into commercial deployment.

Invisible safety zones emerge

“Without this safety certification, our technology doesn’t leave the lab, and it doesn’t reach the factory floor,” Sensory Robotics CEO Chris Edwards says. “Now, it can.”

Industrial robots have long been separated from workers by fences and cages because of the speed and force at which they operate. While collaborative robots have been designed to work closer to humans, many traditional robotic systems still require physical barriers to meet safety requirements.

Sensory Robotics says SR-1 addresses that challenge by continuously tracking people near robotic cells and adjusting robot behavior in real time.

“We designed SR-1 to continuously see people and adjust robot movements in real time,” Edwards says. “This UL certification now allows our technology to be purchased and installed on industrial robots, enabling them to safely operate alongside humans without physical barriers.”

According to the company, the certification process evaluated how the system responds to equipment failures, safety-function disruptions, and potential risks to workers. Such certifications are often required before technologies can be introduced into regulated manufacturing facilities.

Factory robots lose fences

The company argues that eliminating physical cages could free up factory floor space, improve workflow flexibility, and reduce production interruptions. Traditional robotic work cells often require complete shutdowns when workers enter restricted areas, creating downtime that can accumulate across large manufacturing operations.

“The goal is simple,” says chief operating officer Mark Gagas. “Keep people safe without stopping production every time someone enters a workspace. In traditional environments, a safety interruption can require a full robot shutdown and restart, sometimes taking up to 10 minutes per incident. Add that up across a production line, and those delays compound into high operational costs.”

The company says its system is already attracting interest from sectors including automotive, aerospace, logistics, manufacturing, and defense. Sensory Robotics also received a $1.25 million Department of Defense SBIR Phase II grant and additional support through the University of Cincinnati’s innovation ecosystem.

“These combined certifications are the key that unlocks real-world industrial adoption and approval for companies to purchase fenceless human-robot collaboration systems like SR-1,” Gagas says. “It means regulated manufacturing environments can now trust and deploy this technology into their robots with confidence.”

NASA’s Jet Propulsion Laboratory (JPL) has developed a 4-foot-long, four-wheeled prototype designed for high-speed, long-distance autonomous exploration on the Moon and Mars. 

Dubbed Exploration Rover for Navigating Extreme Sloped Terrain (ERNEST), it underwent field tests in the Colorado Desert. The robotic rover successfully traveled 16 miles over 37 hours with minimal human intervention.

“This testing is helping us refine the mobility hardware and autonomy software to navigate extreme distances across a wide range of terrain and lighting conditions anticipated on the Moon,” said Issa Nesnas, a principal technologist at JPL. 

Active suspension system

For the last thirty years, NASA has used a mechanical setup called the “rocker-bogie” suspension system. It is the architecture behind every Mars rover from Sojourner to Perseverance that keeps their wheels grounded.

But it has some problems. Send a passive rover up a loose, sandy slope steeper than 15 degrees, and it will likely get stuck, spin its wheels, or risk a catastrophic tip-over. 

The ERNEST prototype introduces an innovative active suspension system that manages its own weight distribution. 

Equipped with four steerable wheels and a front-powered gimbal, the rover can navigate extreme terrain using unique gaits like squirming, wheel-walking, obstacle-climbing, and even driving sideways. 

Plus, a built-in clutch mechanism allows ERNEST to switch back to a passive mode when efficient, less-demanding driving is needed, blending high-capability mobility with energy conservation.

The 4-foot-long prototype is a testbed designed to advance both robotic autonomy and extreme terrain mobility for future Moon and Mars missions.

ERNEST achieves higher speeds and longer driving distances through its enhanced independent decision-making and a unique ability to lift its individual mesh wheels over obstacles that would trap current Mars rovers.

Ultimately, these technological improvements will inform the design of next-generation rovers capable of exploring previously inaccessible, rough extraterrestrial landscapes.

During recent testing, the vehicle clocked speeds up to 0.6 mph (1 kph) over 37 hours of driving across seven days — an order of magnitude faster than the autonomous navigation speeds of Curiosity and Perseverance. 

“You could do a science road trip across the Moon — or Mars — with this vehicle,” said James Keane, a JPL planetary scientist working on lunar missions.

Preparing for the rough, unknown terrain

Using reinforcement learning, computers ran thousands of hours of virtual simulations over single weekends, punishing the virtual rover with impossible physics scenarios until it learned the optimal way to climb, crawl, and survive.

The software was then loaded directly into the physical rover. Engineers took it out into the desert heat and turned it loose. ERNEST handled the rest. 

The team even ran the rover through the deep, confusing shadows of dusk and dawn. This simulated the extreme lighting conditions future astronauts and robots will face at the Moon’s south pole.

Building on this success, a new project will merge ERNEST’s active suspension control with long-range intelligent navigation. This will allow the rover to plan paths, deciding when to climb surmountable obstacles and when to steer clear of major hazards. 

It will help to prepare future missions for the most rugged landscapes on the Moon and Mars.

💾

Canadian robotics company Sanctuary AI has reported a significant achievement in industrial automation.

The firm said its system achieved a task success rate of more than 99.5 percent during a complex wire-plugging task for a global Tier 1 automotive supplier. 

The operation completed in 2.54 seconds, meeting the customer’s live-production benchmarks and demonstrating high-speed, reliable performance in a manufacturing setting. 

Sanctuary AI said the milestone highlights its strategy of deploying Physical AI on commercial robots to speed industrial adoption and support future industrial humanoids.

Factory automation leap

The achievement reflects Sanctuary AI’s strategy of deploying its Physical AI platform on existing and next-generation industrial robotic systems, rather than waiting for widespread commercialization of humanoid robots. This hardware-agnostic approach enables the company’s AI models to operate across different robotic platforms while accelerating adoption in manufacturing environments.

Sanctuary AI demonstrated high-performance Physical AI capabilities in an industrial manufacturing environment, achieving a task success rate exceeding 99.5 percent in a complex wire-plugging operation for a global Tier 1 automotive supplier. The system completed the task with a cycle time of 2.54 seconds, matching the throughput requirements of the customer’s live production line and meeting established industrial performance benchmarks.

The deployment focused on a plug-insertion process involving flexible wire assemblies moving along a conveyor. Such tasks are particularly challenging for automation because the cables can bend, twist, and shift position dynamically during transport, requiring continuous perception, adaptation, and precise manipulation.

According to the firm, the AI-powered robotic system successfully detected, tracked, aligned, and inserted the connectors while maintaining production-line speed and reliability.

“Manipulating a flexible wire into a moving target on a live conveyor is exactly the kind of contact-rich dexterity problem that has kept tasks like this out of reach for traditional automation. Solving it required models built around performance from day one with reliability, cycle time, and safety measured against real production benchmarks,” said Olivia Norton, co-founder and CTO of Sanctuary AI, in a statement.

AI power industries

Sanctuary AI develops Physical AI systems designed to enable robots to perform complex tasks in real-world industrial environments. Unlike traditional industrial automation, which relies on fixed programming for repetitive operations, Sanctuary AI’s approach combines advanced perception, reasoning, and manipulation capabilities, allowing robots to adapt to changing conditions and handle tasks that require dexterity and decision-making.

The company’s Physical AI platform processes information from cameras, sensors, and robotic actuators to understand its surroundings and execute actions in real time. This enables robots to identify objects, track movement, plan motions, and manipulate components with a high degree of precision. The technology is particularly suited to manufacturing environments where variations in object position, orientation, or movement make conventional automation difficult.

A key feature of Sanctuary AI’s strategy is its hardware-agnostic design. The AI can be deployed on existing industrial robotic systems as well as future robotic platforms, reducing the need for specialized hardware. This allows manufacturers to integrate advanced AI capabilities into current production lines while preserving existing infrastructure investments.

The platform is designed to automate labor-intensive tasks, including assembly, material handling, inspection, and component insertion. By combining real-time perception, motion planning, and adaptive control, Sanctuary AI aims to help factories improve productivity, maintain consistent quality, address labor shortages, and prepare for the next generation of intelligent industrial robots.

“By focusing on a performance-first approach to Physical AI models, we’re delivering value to customers today, on an AI platform that will also scale to the next generation of general-purpose systems,” said Norton, in a statement.

💾

Sen. Bernie Sanders has unveiled a proposal that would give Americans a direct financial stake in the country’s booming artificial intelligence industry. The Vermont independent introduced legislation Thursday that would create a national sovereign wealth fund tied to major AI companies. Under the proposal, the federal government would hold a 50% ownership stake in qualifying AI firms and distribute annual payments to U.S. citizens.

Sanders argues that today’s AI systems rely heavily on content, knowledge, and creative works produced by generations of people. Because AI companies benefit from that collective human output, he says the public should share in the industry’s financial gains. The legislation, called the American AI Sovereign Wealth Fund Act, comes as lawmakers debate how to regulate AI and address concerns about wealth concentration in the technology sector.

Public stake in AI

The proposal would apply to AI companies generating more than $200 million in annual revenue. Those firms would become subject to a system that grants the public a significant ownership interest through a federally managed investment fund. Sanders estimates the fund could reach a value of roughly $7 trillion based on current valuations of leading U.S. AI developers.

The future of AI must not be decided behind closed doors by billionaires seeking to maximize their power and profit.

It should be decided by the American people.

That's why I'm introducing the American AI Sovereign Wealth Fund Act. pic.twitter.com/sbC0YMT90f

— Sen. Bernie Sanders (@SenSanders) June 18, 2026

The senator says the structure would allow Americans to benefit directly from one of the fastest-growing industries in the economy. He projects that citizens could initially receive annual payments of around $1,000. Those payouts could increase as the AI sector expands.

Speaking to reporters, Sanders said the payments would likely grow if AI companies continue generating larger profits. He also suggested future revenue could help support public priorities such as healthcare, education, and housing.

Collective value argument

At the center of the proposal is Sanders’ belief that AI companies derive value from information created by society as a whole. The legislation argues that AI models train on vast amounts of human-created material, including books, journalism, software code, scientific research, artwork, videos, and other forms of knowledge.

Sanders contends that a small group of technology executives now captures most of the economic benefits generated from those resources. Supporters of the measure compare AI to publicly owned natural resources. In that view, companies extracting value from a shared asset should return a portion of that wealth to the public.

Model draws from funds

The bill takes inspiration from sovereign wealth funds already operating across the United States and around the world. States such as Texas and New Mexico have long used investment funds supported by natural resource revenues to finance public programs, particularly education. Dozens of countries, including Norway, China, and the United Arab Emirates, also operate sovereign wealth funds that manage public assets and investments.

If Congress approves the legislation, a seven-member independent commission would oversee the fund. The president would nominate commissioners, while the Senate would confirm them.

The commission would represent public interests in qualifying AI companies and manage the government’s ownership stake. While the proposal faces significant political hurdles, it adds a new dimension to the growing debate over who should benefit from the economic gains created by artificial intelligence.