Xiaomi's release of a massive 38-billion-parameter foundation model is changing the math for synthetic data generation in robotics. While the open-source community gets a powerful new tool, Japan's industrial titans are moving in the opposite direction, officially locking arms with NVIDIA to build a proprietary national physical AI platform. In the background, venture capital is flooding the zone, minting a new $1.1 billion Toyota spinout and backing platforms designed strictly to feed these machines real-world data.
As we noted on Wednesday, Xiaomi has officially open-sourced its 38-billion-parameter Xiaomi-Robotics-U0 foundation model. The key new detail from today's release is its performance: the model reportedly accelerates synthetic training data generation by up to 83 times using its UNIS inference framework, tackling the data scarcity bottleneck head-on. It unifies capabilities like generating interactive 3D scenes from text and transferring trajectories to new scenes, achieving top scores on benchmarks like WorldArena.
Why it matters
The open-sourcing of a large-scale foundation model by a major player like Xiaomi is a significant accelerant for the entire robotics field, especially for open-source developers. By providing a powerful tool to generate diverse and high-quality synthetic data, it drastically lowers the cost and time required for real-world data collection—a primary bottleneck in training capable robots. This could democratize access to advanced robot learning, help close the persistent sim-to-real gap, and speed up the development of more robust policies for everything from industrial manipulators to humanoid robots.
TechTimes emphasizes that U0 addresses the data scarcity problem by leveraging large internet-trained visual models for embodied generation and transfer. TechNode highlights the model's ability to generate environments from text and adapt existing trajectories to new scenes. OpenSourceForU notes the model's top score on the WorldArena benchmark and the 83x speed-up from its UNIS inference acceleration framework.
Researchers at Georgia Tech have developed a more efficient machine learning framework called 'Learn to Teach' for training bipedal robots. The method, which was published in Science Robotics, simultaneously trains a 'student' policy to control the robot and a 'teacher' policy that learns to create the most effective training curriculum for the student. This approach reportedly enabled a full-sized humanoid robot to successfully navigate challenging, varied terrains like gravel, grass, and slopes with significantly less computational cost and training time compared to traditional reinforcement learning methods.
Why it matters
This research addresses a major bottleneck in robotics: the immense time and computational power required to train robots to be robust in unpredictable real-world environments. By making the training process itself more intelligent and efficient, this 'Learn to Teach' framework could dramatically accelerate the development cycle for legged robots. This makes it more feasible and cost-effective to create versatile robots for applications like search and rescue, logistics, and exploration where adaptability to unstructured terrain is critical.
The Georgia Tech College of Engineering release states the method allows a two-legged humanoid robot to successfully navigate various challenging terrains. Interesting Engineering notes the robot trained with this framework outperformed even the manufacturer's own controller on slippery surfaces. Nordlandfun emphasizes that the framework accelerates the process and improves the robot's ability to generalize to new environments.
Researchers at the Korea Advanced Institute of Science and Technology (KAIST) have detailed the APT-RL control technology behind the 'KAIST Hound' quadruped we've been tracking. The framework allows the robot to autonomously switch between different gaits—learning walking, running, and jumping locomotion skills to analyze upcoming terrain and select the best movement strategy in real-time. Recent outdoor tests validated its stability across grass, paved trails, and athletic fields.
Why it matters
This research moves beyond single-gait locomotion to a more versatile, adaptive system that better mimics how animals move. The ability for a robot to autonomously choose the right 'tool' for the job—whether that's a slow, careful walk or a dynamic jump—is crucial for navigating the unstructured and unpredictable environments found in the real world. This has direct applications for robots in disaster response, defense, and industrial inspection, where terrain cannot be guaranteed.
Techxplore.com reports that the new learning-based control technology, APT-RL, allows the robot to autonomously select and switch between various locomotion skills. This enables the robot to move stably and quickly across complex outdoor terrains by adapting its gait in real time, overcoming previous limitations.
Confirming a partnership we've been tracking, Japan's Ministry of Economy, Trade and Industry has announced that 22 of the country's leading robotics and manufacturing firms are forming a coalition with NVIDIA to build a national physical AI infrastructure. The group, which includes industrial heavyweights like FANUC, Yaskawa Electric, Kawasaki Heavy Industries, and Fujitsu, will use NVIDIA's full stack of AI and robotics platforms—including Isaac, Metropolis, and Jetson—to develop open, multimodal foundation models. As part of the initiative, NVIDIA also launched Cosmos 3 Edge, a new 4-billion-parameter world model designed to run on-device for real-time vision reasoning and robot control.
Why it matters
This official, large-scale coalition marks a major strategic commitment by Japan to establish a leadership position in the next wave of industrial automation. By creating a unified, open platform for physical AI, the nation aims to leverage its deep manufacturing expertise to accelerate the deployment of intelligent machines. For the robotics industry, this centralized effort could streamline development, establish common standards, and create a powerful ecosystem for training and deploying robots, moving the frontier of physical AI from the data center to the factory floor.
TechTimes highlighted that 22 Japanese firms plan to join the Cosmos Coalition to build open world models for physical AI. SiliconANGLE noted this move signifies NVIDIA's aggressive strategy to dominate the physical AI market beyond data centers. ADT Magazine emphasized the partnership's goal to enable real-time, autonomous robot operation by pushing advanced AI to edge devices, reducing latency and cloud reliance.
Following the Jetson kit price cuts we tracked last week, NVIDIA is continuing its edge hardware push with two new additions to the Jetson Thor family: the T3000 and T2000 modules. Based on the Blackwell architecture, the compact T3000 offers 865 FP4 TFLOPS for humanoids and mobile robots, while the T2000 provides 400 FP4 TFLOPS for edge AI applications. NVIDIA also announced new 'Jetson agent skills' software to optimize memory usage against rising DRAM prices, with both modules slated for Q1 2027.
Why it matters
These new modules are a clear move by NVIDIA to democratize high-performance AI and push its platform into mass-market robotics. By offering a spectrum of compute options, NVIDIA lowers the barrier to entry for developers and companies building smaller, more power-constrained, or cost-sensitive robots. For a robotics entrepreneur, this expands the hardware options for on-device inference, making it more feasible to move from pilot projects to large-scale commercial deployments without being forced into the highest-cost hardware tier.
Glitchwire reports the new modules and software are a direct response to escalating memory costs and supply chain constraints, enabling more efficient hardware use. Wccftech highlights the extension of the Blackwell GPU architecture to mainstream robotics. IoT Tech News notes this could significantly accelerate the transition from pilot fleets to widespread commercial deployment. Ecosystem partners like Advantech and Connect Tech immediately announced support for the new modules.
The 2026 World Artificial Intelligence Conference (WAIC) in Shanghai is heavily focused on advancements in AI and humanoid robotics, underscoring China's strategic priorities. Unitree Robotics unveiled its mass-produced transformable robot, the GD01, which can shift between a quadruped and a wheeled humanoid form capable of carrying a human. Other exhibitors, like Pudu Robotics, showcased industrial semi-humanoid robots, while AGILINK and Agibot displayed dexterous manipulation capabilities. A key theme of the event is the progress of China's domestic supply chain for core components. In his opening address, President Xi Jinping emphasized the need for international cooperation in AI.
Why it matters
WAIC 2026 serves as a clear demonstration of China's ambition to lead in applied robotics, moving from R&D to mass-producible, commercially viable products. The sheer variety of robots on display—from transformable human-carriers to industrial humanoids—and the explicit focus on a robust domestic supply chain signal a concerted national effort to industrialize the sector. The simultaneous call for international cooperation suggests China wants to both compete and help shape global AI governance.
SBS News reported on Unitree's transformable GD01 robot and President Xi's call for cooperation. PR Newswire highlighted Pudu Robotics' 'One Brain, Multiple Embodiments' architecture and its new industrial semi-humanoid. The Manila Times noted that over 300 new AI products are making their debut, with a strong focus on practical applications from companies like SenseTime and BrainCo.
Fresh off the $200 million pre-IPO funding round we tracked earlier this week, LimX Dynamics has released a new demonstration of its full-size Oli humanoid autonomously performing household chores in a real home setting. The robot folded clothes, wiped a table, and organized items without teleoperation, powered by its proprietary COSA 0.5 'brain' system—which notably uses a modular architecture rather than a single large end-to-end model.
Why it matters
This demonstration places LimX alongside Figure in showcasing a humanoid capable of complex, multi-step autonomous operations in an unstructured environment. Their explicit choice of a modular AI architecture, contrasting with the monolithic model approach pursued by some competitors, presents an alternative thesis for achieving capable physical AI. For the robotics community, this provides another key data point on which software strategies are proving most effective for tackling real-world manipulation tasks.
According to Robotic Firms, this achievement positions LimX Dynamics alongside Figure AI as a leader in long-horizon autonomous manipulation in a home environment. The company's technical path, which is distinct from a single large model approach, suggests a modular architecture might be more effective for complex physical interaction. The demonstration highlights progress towards developing physical AI for consumer and service markets.
Humanoid AI, the London-based startup we noted was in talks for a $200 million funding round, has detailed its commercial strategy. The company is prioritizing wheeled versions of its HMND 01 robot for initial industrial deployments, citing easier certification and greater factory logistics efficiency. Built on a four-layer 'brain' architecture, the firm reports it has already secured deployment deals with European giants like Schaeffler, Bosch, and Siemens, aiming for a 100,000-unit production capacity by 2031.
Why it matters
Humanoid AI's focus on wheeled platforms for its initial commercial push is a notable strategic choice that contrasts with the bipedal-first approach of many competitors. By targeting achievable certifications and immediate industrial needs, the company is opting for a faster path to revenue and large-scale data collection. This highlights a potential segmentation in the humanoid market, with some players focusing on bipedal dexterity while others prioritize near-term logistics automation, even if it means sacrificing a fully human-like form factor.
According to The Financial News 24/7, the company is focusing on wheeled versions to ease certification and improve efficiency for industrial applications. The strategy appears to be paying off, with deployment deals already in place with major German industrial partners. This pragmatic approach positions the company as a strong European contender in the global robotics race.
Chinese electric vehicle maker XPeng is accelerating its robotics ambitions, targeting a production rate of over 1,000 'Iron' humanoid robots per month by the end of 2026. The company plans a global rollout for the robot in 2027, beginning with deployments as sales assistants in its own showrooms across China in the first quarter of that year. The move leverages XPeng's existing expertise in AI and autonomous systems from its automotive division.
Why it matters
XPeng's aggressive production target represents another automaker making a serious, large-scale push into humanoid robotics, following similar announcements from Tesla, Hyundai, and others. By setting a concrete production goal and deployment timeline, XPeng is signaling its intent to move beyond R&D and establish a commercial presence in the service robotics market. Using its own showrooms as the initial deployment site provides a controlled environment to gather data and refine the robot's capabilities before a wider launch.
Interesting Engineering reports that the robots are slated for a global rollout in 2027, with an initial deployment as sales assistants in XPeng's Chinese showrooms. The move is seen as a strategic expansion for the EV maker, leveraging its AI and autonomous driving expertise to enter the physical AI market.
Ace Robotics has unveiled and open-sourced Kairos 3.0, a 'world model' for embodied AI that it claims achieves state-of-the-art (SOTA) results on four major international benchmarks. The architecture is described as a 'natively unified multimodal understanding-generation-prediction' system. Notably, the company is also releasing a lightweight 4-billion-parameter version of the model that it says runs 72 times faster than competing models, making it the first open-source world model capable of real-time deployment on edge hardware. The release includes code and datasets.
Why it matters
While many foundation models for robotics require significant cloud computing resources, Ace Robotics' release of a high-performance model specifically optimized for the edge is a significant contribution to the open-source community. Providing a production-ready 'AI brain' that can run locally on robots lowers a major barrier to deployment. This could empower smaller companies and individual developers to build and test sophisticated autonomous behaviors without relying on expensive cloud infrastructure, accelerating the move from lab-based demos to real-world applications.
According to AIjourn, the Kairos 3.0 model is designed for commercial deployment across various robot categories, including inspection, logistics, and home automation. The company's goal is to push embodied AI from the lab to large-scale commercial use. The open-sourcing of both the model and its datasets is intended to empower the developer community.
As we noted recently, Munich-based Microagi secured a massive $55 million seed round. The 10-month-old company, founded by ex-Formula One engineers and an Alan Turing researcher, has now detailed that the capital will fund 'Atlas'—a new platform designed to capture operational data from factory workers. Rather than building robots or foundation models, Atlas focuses on fine-tuning existing AI models for specific industrial tasks to solve the 'last mile' reliability problem in real-world settings.
Why it matters
This massive seed round for a company focused purely on the data layer—not hardware or models—is a powerful market signal. It underscores that a key bottleneck to widespread robot deployment is the lack of high-quality, task-specific training data. Microagi's approach of creating a 'data-as-a-service' layer to make existing robots smarter, rather than building new ones, represents a significant business model innovation. For robotics entrepreneurs, it highlights a massive opportunity in the enabling infrastructure around AI, not just in the robots themselves.
TechFundingNews notes the company was founded by two F1 engineers, a WhatsApp entrepreneur, and an Alan Turing researcher. SiliconANGLE clarifies that Microagi's platform is designed to teach factory robots how to work by fine-tuning existing AI models. Semafor points out the startup's globalized approach, combining Chinese hardware with German software and U.S. data, highlighting the interconnected nature of the modern robotics supply chain.
The $300 million seed round for a Toyota spinout that we noted earlier this week belongs to Walden Robotics, which has now officially emerged from stealth. Built by veterans from Toyota Research Institute, MIT, and Stanford, the company is developing a general-purpose robotics platform for industrial applications. Its wheeled, upper-body humanoid robots are already deployed in Toyota plants, utilizing 'Large Behavior Models' and Diffusion Policy for continuous imitation learning rather than explicit programming.
Why it matters
Walden's massive seed round and immediate unicorn status underscore immense investor confidence in a more pragmatic approach to industrial robotics. Instead of pursuing general-purpose autonomy, Walden is focused on a specific vertical (manufacturing) with a proven learning model (imitation learning) and an immediate deployment partner (Toyota). This go-to-market strategy could significantly de-risk the path to commercialization and sets a new standard for how AI-native robotics companies can achieve rapid scale by solving concrete industrial problems from day one.
Tech Funding News highlights the founding team's pedigree from TRI, MIT, Stanford, and Amazon. Interesting Engineering notes the platform's use of Large Behavior Models and Diffusion Policy enables continuous learning from real-world experience. The Boston Globe adds that the company was co-founded by MIT professor Russ Tedrake and is using a 'robots as a service' model.
Monumental, a Dutch robotics company that develops autonomous bricklaying robots, has raised $32 million in a Series B funding round led by Khosla Ventures. The company's robots, which are small, electric-powered, and designed to work alongside human masons on construction sites, use AI and computer vision to build walls. Monumental operates as an autonomous subcontractor, providing both the robots and the service, rather than selling the hardware. The funding will be used for European expansion and entry into the U.S. market.
Why it matters
Monumental's funding and business model highlight a practical approach to construction automation. By focusing on a specific, labor-intensive task (bricklaying) and offering a 'robotics-as-a-service' model, the company lowers the adoption barrier for construction firms that may be hesitant to make large capital investments in new technology. This strategy of augmenting, rather than replacing, human workers could accelerate the integration of robotics into the historically slow-to-adopt construction industry.
The AI Insider reports the investment will fuel the company's expansion across Europe and entry into the U.S. market. The company's unique business model of acting as an autonomous subcontractor, rather than just selling machines, is seen as a key factor in its potential for rapid adoption.
Switzerland-based Mimic Robotics AG has launched the Mimic hand M1.0, a specialized robotic hand designed to emulate human dexterity for industrial automation. The hand features bidirectional pulley-guided tendons and tactile fingertip sensors. Crucially, it is trained using the companion 'mimic wearable U1,' an exoskeleton device that captures human hand data directly. This allows the robot to learn complex manipulation tasks through human demonstration rather than simulation or teleoperation.
Why it matters
Mimic's platform directly tackles one of the biggest challenges in robotics: collecting high-quality, real-world data for training manipulation. By using a wearable to capture human motion, they can create a tight feedback loop for imitation learning, potentially bypassing the sim-to-real gap and accelerating the automation of tasks that have historically been too complex or delicate for robots. This 'full-stack' approach of building the hand, the wearable, and the learning platform in-house represents a vertically-integrated strategy to solve dexterous manipulation.
SiliconANGLE highlights the hand's ability to train using both human video pretraining and physical data from wearables. Humanoid Robotics Technology notes the company's full-stack approach, with in-house design and manufacturing, aims to achieve general-purpose dexterous manipulation. The goal is to automate manual tasks that have previously been uneconomical or impossible to robotize.
At its 'Smart Manufacturing' exhibition, Foxconn debuted a new suite of industrial AI robots, including humanoid robots for production lines, wafer handling robots, and remote-controlled dexterous hands. Developed in partnership with NVIDIA, the robots are trained using AI in virtual factory environments (digital twins) before deployment. This allows for rapid optimization and enables the robots to perform high-precision tasks like component assembly and wafer transport with greater adaptability.
Why it matters
As one of the world's largest contract manufacturers, Foxconn's deep investment in AI-driven robotics signals a major shift in global production. The use of NVIDIA's simulation platforms to train robots before they hit the factory floor can dramatically reduce commissioning time and costs. This move not only enhances Foxconn's own efficiency but also serves as a powerful proof-of-concept for the entire manufacturing sector on how to scale advanced, flexible automation.
Boardor.com reports that the robots leverage AI for training in virtual environments, enabling them to handle complex tasks with high efficiency. The collaboration with NVIDIA is key to this digital twin strategy. The goal is to enhance precision, enable continuous 24/7 operation, and ultimately reduce production costs across its vast manufacturing operations.
Updating a story from earlier in the week, Sunday Robotics has now provided more detail on its home robot, Memo. The company claims its new AI model, ACT-2, can fold laundry it has never seen before in unfamiliar home environments with over 99% success. Sunday also proposed a new robotics industry benchmark called a 'Solve,' defined as a task a robot can perform with at least 99% reliability across 100 different households. The company plans to launch a beta program for Memo this fall.
Why it matters
Laundry folding has long been a benchmark challenge for robotic manipulation due to the complexity of handling deformable objects. If independently verified, a 99% success rate in novel environments would represent a major breakthrough for practical home robotics. The proposed 'Solve' standard is also significant, as it attempts to introduce a rigorous, real-world metric for reliability in a field often characterized by impressive but brittle demos. This could bring much-needed clarity for consumers and developers alike.
Business Insider reports that the new ACT-2 model demonstrated the high success rate with untrained garments. Dnyuz.com adds that the company is proposing the 'Solve' as a new industry standard to measure progress reliably and plans a beta program for the Memo robot this fall.
Orthopedic device giant Stryker has announced the full U.S. commercial launch of its Mako RPS (Robotic Power System) for total knee replacement. The device is a handheld robotic system that integrates the company's Mako smart robotics technology and 3D planning with a familiar power tool form factor. The system uses sensors to provide surgeons with real-time feedback and active adjustment technology to help execute precise bone cuts according to a pre-operative plan.
Why it matters
The Mako RPS represents a significant expansion of the surgical robotics market into a new category of handheld, intelligent tools. Unlike large, standalone robotic systems, this platform integrates robotic precision directly into the surgeon's hands, potentially lowering the barrier to adoption in terms of cost, operating room footprint, and workflow disruption. This could accelerate the spread of robotic-assisted techniques beyond major hospitals into ambulatory surgery centers.
MassDevice reports that the launch introduces a new category of handheld orthopedic robotics, making the technology accessible to a broader market segment. StockTitan adds that the system is compatible with Stryker’s existing Triathlon Total Knee System and Q Guidance System, allowing it to fit into established workflows.
Surgical robotics leader Intuitive Surgical reported strong second-quarter results, beating Wall Street expectations. Worldwide procedures using its da Vinci and Ion robots grew by approximately 16% compared to the previous year, and the company shipped 444 da Vinci systems. Despite the positive results, the company's stock dipped, which analysts attribute to broader market caution, regulatory delays for its next-gen da Vinci Edge platform, and increasing competition from other surgical robotics firms.
Why it matters
Intuitive's continued growth in procedure volume demonstrates the ongoing, widespread adoption of robotic-assisted surgery and the company's entrenched market position. However, the market's reaction and analyst commentary reveal the pressures of a maturing industry. The combination of regulatory hurdles, emerging competitors, and investor expectations for near-perfect growth creates a more challenging landscape than in previous years, signaling that even the dominant incumbent is not immune to market forces.
MassDevice reported that profits reached $818.1 million on sales of $2.89 billion, exceeding analyst forecasts. Despite this, AktienSensor notes the share price declined due to factors including regulatory delays for the da Vinci Edge and aggressive pricing from competitors, highlighting a more complex and competitive market environment.
A research collaboration between the Korea Advanced Institute of Science and Technology (KAIST) and Stanford University has developed a soft robotic technology that allows clothing to dress a person automatically. The system uses air-powered, vine-like robots inspired by climbing ivy, which are embedded within the fabric. When activated, these 'vines' grow and guide the garment onto the wearer's body in about 10 seconds, even accommodating movement. The technology was detailed in the journal Science Robotics.
Why it matters
This is a novel and practical application of soft robotics that moves beyond grippers and manipulators into wearable technology with a direct human-assist function. The most immediate impact is for assistive technology, offering greater independence to the elderly and individuals with mobility impairments. However, it also has significant potential in industrial and emergency settings, such as for the hands-free donning of cleanroom suits or complex protective gear for first responders, improving both safety and efficiency.
Reuters demonstrated the technology with a video showing a life jacket and pants putting themselves on a mannequin. Asharq Al-Awsat notes the system can dress a person in about 10 seconds. EconoTimes emphasizes its application for workers in semiconductor cleanrooms or emergency services who need to suit up quickly and without contamination.
National AI Strategies Solidify Around Physical AI Infrastructure Following up on a partnership announced earlier this week, Japan's Ministry of Economy, Trade and Industry, along with 22 industrial giants like FANUC and Yaskawa, are officially forming a coalition with NVIDIA to build a national AI infrastructure focused on physical AI. This mirrors China's strategy of mass deployments to gather real-world data and highlights a global trend of government-backed initiatives to secure a lead in robotics and automation.
Venture Capital Targets Both Robotics Hardware and the Underlying Data Layer A wave of significant funding rounds shows investors are betting on both ends of the robotics stack. Walden Robotics, a Toyota spinout, raised a $300M seed round for its hardware, while Munich-based Microagi secured Germany's largest-ever seed round ($55M) to build a platform that collects factory data to train robots. This dual focus suggests the market sees value not just in the machines themselves, but in the specialized data and software required to make them intelligent.
Major Tech Players Open-Source Large-Scale Robotics Models Xiaomi has open-sourced Xiaomi-Robotics-U0, a massive 38-billion-parameter foundation model designed to dramatically accelerate synthetic data generation for robot training. This follows similar open-source releases from Ant Group and Hugging Face, indicating a strategic push by major technology companies to establish their architectures as industry standards and accelerate development across the ecosystem.
Soft Robotics Advances with Self-Dressing Clothes and Durable Materials Researchers from KAIST and Stanford have developed 'self-dressing' clothing using air-powered, vine-like soft robots embedded in the fabric. This practical application of soft robotics is complemented by materials science breakthroughs, such as a new 3D-printable elastomer from EPFL that is highly resistant to fracture and fatigue, addressing a key durability challenge for flexible robots.
Handheld Surgical Robotics Gains Momentum Stryker has fully launched its Mako RPS, a handheld robotic power system for knee replacements, expanding its surgical portfolio. The launch signifies a trend toward more compact, accessible robotic systems that integrate advanced planning and precision into familiar, handheld form factors, potentially broadening the adoption of robotic assistance in operating rooms.
What to Expect
2026-07-22—Tesla Q2 2026 earnings call, where shareholders are expected to press for updates on Robotaxi and Optimus progress.
2026-08-05—A webinar on 'How Motors, Drives, and Sensors are Advancing Humanoid Robots,' hosted by Tech Briefs, featuring experts from EnduX, maxon, and Bosch.
2026-09-09—IMTS 2026, the International Manufacturing Technology Show, begins in Chicago, with a major focus on industrial AI and automation.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
642
📖
Read in full
Every article opened, read, and evaluated
242
⭐
Published today
Ranked by importance and verified across sources
19
— The Robot Beat
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste