We're seeing a sharp divergence in how the robotics industry approaches hardware and control today. A cascade of massive funding rounds is flowing into startups across the US and Europe, targeting everything from AI-native industrial arms to data center maintenance bots. At the same time, a wave of open-source software releases is aiming to democratize advanced control, arriving just as new edge hardware hits the market to run those models locally.
South Korean conglomerate Hanwha Group has formally launched Hanwha Machinery & Services Holdings (Hanwha M&S), a new entity that consolidates 57 affiliates under a single 'physical AI' mandate. The move combines Hanwha Vision (AI cameras), Hanwha Semitech (semiconductor equipment), Hanwha Robotics (cobots), and Hanwha Momentum (factory automation) into a vertically integrated stack. The structure also includes Hanwha's lifestyle businesses, such as retail and hotels, which will serve as a captive testbed for deploying and refining its new robotic systems.
Why it matters
Hanwha's restructuring is a massive bet on vertical integration as the path to solving the 'sim-to-real gap'—one of the hardest problems in robotics. By controlling everything from chip manufacturing equipment to the end-user environment where robots are deployed, Hanwha can create a powerful feedback loop for data collection and model refinement. For entrepreneurs in the robotics space, this move provides a compelling, if difficult to replicate, blueprint for how a major industrial player is tackling the deployment challenge at scale, and it raises the competitive stakes for companies focused on only one part of the physical AI stack.
The strategy aims to overcome the data scarcity and domain shift problems that plague robot AI development by creating a closed-loop system. The integration of its own retail and hospitality businesses as a real-world lab could give Hanwha a significant advantage in gathering the proprietary operational data needed to train robust, generalizable AI models for service and industrial applications.
Apptronik has launched Apollo 2, its next-generation AI-powered humanoid robot platform designed specifically for continuous real-world learning in industrial settings. Available in both bipedal and wheeled versions, Apollo 2 is engineered to collect vast amounts of operational data from its deployments. This data will be used to train and refine Google DeepMind's Gemini Robotics foundation models, creating a direct feedback loop from physical interaction to AI model improvement.
Why it matters
This launch formalizes a critical link between advanced humanoid hardware and leading AI models. Instead of one-off demos, Apollo 2 is being positioned as a data-gathering fleet to solve the core bottleneck in embodied AI: the scarcity of high-quality, real-world interaction data. This strategy directly addresses the 'sim-to-real' gap by grounding AI development in the complexities of physical environments, aiming to accelerate the creation of more reliable and adaptable robots for practical industrial use.
The move reinforces the idea that the next phase of progress in robotics will be driven by data from deployed systems, not just simulation. By partnering with Google DeepMind, Apptronik is betting that providing the data pipeline to a leading foundation model will be a faster path to commercial viability than developing a full AI stack in-house. This contrasts with vertically integrated approaches like Tesla's and now Hanwha's.
We've been tracking NVIDIA's rollout of both the Jetson Thor architecture and the Cosmos 3 'omnimodel' over the past few weeks. Now, NVIDIA has officially launched the T3000 and T2000 compact Thor modules, alongside Cosmos 3 Edge—a 4-billion-parameter open world model optimized specifically for these devices. The combination allows robots to perform local reasoning and achieve 15 Hz control without cloud connectivity.
Why it matters
This dual release is a significant step toward commoditizing the hardware and software required for sophisticated on-device AI. By making high-performance compute more accessible and providing a capable, open model to run on it, NVIDIA is lowering the barrier to entry for developing and deploying intelligent autonomous systems. This could accelerate the transition of physical AI from research labs and high-end applications into more mainstream commercial and consumer robots.
NVIDIA's strategy appears to be creating an entire ecosystem, from hardware to foundational models, to capture the edge robotics market. The open-source release of Cosmos 3 Edge on Hugging Face is a key move to foster a developer community and drive adoption of its Jetson platform, directly competing with alternative hardware from companies like Qualcomm, AMD, and a growing number of specialized chip startups.
Indian AI startup Sarvam AI is raising $74 million in a Series B extension led by a $25 million investment from NVIDIA. The new funding round will push the company's valuation to approximately $1.51 billion. Sarvam AI focuses on developing large language models and the corresponding inference infrastructure specifically tailored for Indian languages and cultural contexts.
Why it matters
NVIDIA's investment in Sarvam AI highlights a key strategic trend: the move towards specialized, regional AI models. While general-purpose models dominate headlines, this deal shows the significant commercial and strategic value in building AI tailored to specific linguistic and cultural markets. For entrepreneurs, this signals a major opportunity in developing niche AI solutions that address needs not fully met by global models, and demonstrates that such ventures can attract backing from top-tier strategic investors.
This funding is not just about language; it's about building a full-stack AI ecosystem for India. By backing Sarvam, NVIDIA gains a strategic foothold in one of the world's largest and fastest-growing digital economies, fostering a partner that can drive demand for its hardware while developing locally-relevant applications across various industries.
Following Figure AI's recent deployment of the Figure 03 at BMW and its autonomous 8-hour shift demonstrations, the company has released a new video showing the humanoid autonomously climbing a ladder. Figure states this complex benchmark task is enabled by its upgraded Helix System 0 AI model, which integrates visual perception with whole-body motion control.
Why it matters
Ladder climbing is a benchmark task that requires a sophisticated combination of balance, coordination, precise foot and hand placement, and sequential task planning. Successfully automating this demonstrates a significant leap in locomotion and real-world utility for humanoids. While still a demonstration, it moves the technology closer to being able to navigate complex, multi-level human environments like construction sites or warehouses, expanding the potential scope of tasks these robots can perform.
This achievement showcases the rapid progress in embodied AI, particularly in policies that can control the entire robot's body in a coordinated way to perform difficult tasks. It's a clear indicator that the focus is shifting from basic walking to more advanced mobility that is essential for practical applications in unstructured settings.
Following the 13,000 pre-orders and ethical debates we tracked around UBTech's U1 companion robot, the company has officially begun deliveries in China. Early adopters are paying 169,800 yuan (approximately $23,680)—notably lower than the $135,000 high-end price point previously floated. Initial reports indicate buyers are valuing the 88-degree-of-freedom robot primarily for its conversational AI rather than utility for household chores.
Why it matters
This is a critical real-world test for the consumer humanoid market. The high price point and the emphasis on companionship over labor test the market's willingness to adopt expensive robots for social interaction. For the consumer robotics industry, the success or failure of the U1 will provide invaluable data on the actual demand for non-utilitarian humanoids and will influence the product strategies of competitors. The current framing suggests that lifelike conversation may reach the home before dependable domestic labor.
Some see this as a pivotal moment, validating the market for companion robots, especially given China's aging population. Others are more skeptical, viewing it as a niche product for wealthy early adopters and questioning whether the technology is mature enough to provide meaningful, long-term companionship. The company itself has reportedly narrowed its initial broad claims about elder care and mental support, highlighting the difficulty of achieving genuine empathy in a machine.
Robbyant, Ant Group's robotics unit, has open-sourced LingBot-VLA 2.0, a vision-language-action (VLA) model designed for cross-morphology training. The model is built to learn from diverse real-world data and generalize its skills across different types of robots, including quadrupeds and robotic arms. Robbyant states the model uses an autoregressive architecture for dynamic world modeling and real-time execution, designed natively for the physical world rather than being fine-tuned from digital content.
Why it matters
Open-sourcing a model capable of working across different robot designs addresses a major pain point in robotics: the high cost and effort of retraining AI for each new piece of hardware. This could significantly accelerate the adoption of embodied AI by allowing developers to leverage a single, powerful model for various applications. For the open-source community, this provides a powerful new tool for building adaptable robots and fosters a more collaborative approach to solving general-purpose robotics.
This move is a strategic play to establish LingBot as a foundational platform in the robotics ecosystem, similar to what open-source LLMs have done for text-based AI. By enabling one model to control different 'bodies,' it could drastically reduce deployment costs and encourage more experimentation with novel robot hardware.
Following its recent launch of the open-source Cyclo Intelligence platform, South Korean firm Robotis is taking a further step by fully open-sourcing the software for its 'AI Sapiens' humanoid. The release includes the company's core control code, 'Sim2Real' tools, and reinforcement learning workflows, aiming to establish its platform as a standard against growing Chinese competition.
Why it matters
This is a significant contribution to the open-source robotics community. By releasing not just high-level code but the entire stack, including critical sim-to-real tools, Robotis is dramatically lowering the barrier to entry for advanced humanoid research and development. This could foster a wave of innovation, allowing startups, academics, and hobbyists to build upon a proven, high-performance platform instead of starting from scratch, potentially accelerating progress across the entire field.
This move can be seen as a strategic effort to build a wide developer ecosystem around the 'AI Sapiens' platform, hoping that community adoption will make it a de facto standard. It's a classic open-source playbook applied to the complex world of humanoid hardware, betting that a large, active community can out-innovate closed, proprietary systems.
Researchers have introduced N_0-VTLA, a new vision-tactile-language-action (VTLA) foundation model that incorporates tactile feedback as a primary input for robot control. The model aims to address a key gap in purely vision-based robot policies, which often struggle with tasks requiring contact awareness. It was pretrained on a large-scale visuo-tactile dataset and uses an offline reinforcement learning method called ALTER to continuously improve its policy from data gathered during deployment.
Why it matters
The ability to 'feel' is critical for dexterous manipulation, allowing robots to handle delicate objects, detect slip, and apply precise force. By integrating tactile data at a foundational level, N_0-VTLA represents a significant step towards more robust and reliable physical AI. The inclusion of an offline RL method for post-deployment improvement is also crucial, as it provides a practical path for robot capabilities to evolve and adapt based on real-world experience, rather than remaining static after training.
This work challenges the current vision-centric paradigm in robot learning. While vision is essential for understanding the scene, touch provides the ground truth for interaction. The researchers argue that a model that fuses both modalities from the start will fundamentally outperform those that rely on vision alone for contact-rich tasks.
Newly published patent filings from Tesla for its Optimus humanoid robot detail key mechanical innovations in its hand and knee joints. The hand design focuses on a novel cable routing system to eliminate 'crosstalk' between finger tendons, enabling more precise and independent finger actuation. The knee joint utilizes a four-bar linkage mechanism designed to provide energy-efficient leverage throughout its range of motion.
Why it matters
These patents underscore a critical trend in modern robotics: solving complex problems with 'smarter mechanics' rather than relying solely on brute-force computation and actuation. The focus on eliminating crosstalk and maximizing energy efficiency at the component level is essential for creating practical, reliable humanoids that can operate for extended periods. For robotics engineers, this highlights the continued importance of fundamental mechanical design in the age of AI.
This is a departure from the 'bigger motors, faster processors' approach. Tesla is investing heavily in the underlying mechanical engineering to make control problems easier to solve in software. The four-bar linkage in the knee, for instance, is a classic mechanism that can create more favorable torque profiles, reducing the load on the motors and saving power.
German cognitive robotics company Neura Robotics has reportedly secured up to $1.4 billion in a new funding round with participation from major tech firms including Nvidia, Amazon, and Qualcomm. The massive investment has catapulted the company's valuation to approximately $7 billion, positioning it as a leading European contender in the global race to build physical AI and humanoid robots.
Why it matters
This is one of the largest funding rounds for a European robotics company to date, signaling that the continent is a serious player in the capital-intensive humanoid race, not just Silicon Valley. The backing from a strategic consortium of chipmakers and cloud providers underscores the industry's shift toward robots that can learn and interact in real-world environments. For the robotics market, Neura's focus on a cognitive AI platform first, and hardware second, represents a significant bet on the 'brains' being the key differentiator.
The investment validates Neura's platform-centric approach, where a common AI brain can be deployed across different robotic form factors, including their MAiRA cognitive cobots and the 4NE-1 humanoid. This contrasts with hardware-first companies and suggests investors see immense value in the underlying software that enables robotic intelligence.
Beijing-based robotics startup PokeBot has closed its Pre-A funding round, raising over $100 million just four months after its incorporation. The company is focused on robot manipulation, which many in the field consider the primary remaining bottleneck for general-purpose robots. This is PokeBot's second nine-figure funding round in three months, co-led by Shunwei Capital and Matrix Partners.
Why it matters
Such a massive early-stage investment into a company focused on a single, difficult problem highlights intense investor interest in solving the core challenges of physical interaction. PokeBot's reported focus on real-robot reinforcement learning, as opposed to sim-to-real, represents a significant architectural bet. Success with this approach could provide a new template for training robots that is less dependent on the fidelity of simulations, potentially accelerating progress in dexterous manipulation.
The funding signals that investors are willing to make large, concentrated bets on teams tackling foundational robotics problems. While simulation-heavy approaches are common, PokeBot's strategy suggests a belief that the nuances of real-world physics and contact dynamics are too complex to simulate perfectly, and that learning directly on hardware is the faster path to success, despite the higher initial cost and complexity.
Zurich-based Exclaim Robotics has come out of stealth mode, securing $4.95 million in funding to develop specialized robots for routine repair and maintenance tasks inside large-scale data centers. The company is led by veteran roboticist Helen Oleynikova, a former researcher at Google and NVIDIA.
Why it matters
This funding highlights a growing trend towards highly specialized robotics targeting niche, high-value industrial problems. Data centers are critical, complex, and increasingly vast infrastructures where automated maintenance can significantly improve uptime and efficiency. Developing robots for this specific environment represents a practical, commercially-focused application of robotics that could provide a faster path to profitability than more general-purpose solutions.
As AI workloads drive the explosive growth of data centers, the operational challenges of maintaining them also scale. Automating tasks like swapping server racks, replacing failed components, and managing cabling is a logical next step. Exclaim's focused approach is a bet that solving a specific, urgent pain point for a booming industry is a more direct path to market than building a generalist robot.
FieldAI, a startup developing AI for robots operating in unpredictable real-world environments like mines and construction sites, has reached $100 million in revenue. The company has achieved a $2 billion valuation by focusing on 'Field Foundation Models' that rely on physics and probability to enable risk-aware AI, a different approach from data-heavy models trained on large, static datasets.
Why it matters
FieldAI's commercial success validates an alternative approach to robot AI that is less dependent on massive, pre-existing datasets. By building models grounded in physics, the company can deploy robots in novel, unstructured environments where collecting training data is impractical or impossible. This revenue milestone proves there is a significant market for robust, adaptable robotic solutions in heavy industry, a sector often overlooked in the hype around consumer and logistics robots.
The company's success challenges the notion that robotics is a 'data-first' problem. Instead, they argue it's a 'physics-first' problem, where understanding the real-world constraints and uncertainties is paramount. This methodology allows them to achieve reliability in environments where traditional machine learning approaches might fail due to a lack of representative training data.
The U.S. FDA has granted 510(k) clearance to the Zeta TMS Robotic System. This system provides real-time, neuronavigated positioning for Transcranial Magnetic Stimulation (TMS) coils with submillimeter accuracy. TMS is a non-invasive procedure used to treat various psychiatric conditions, including treatment-resistant depression and OCD, by using magnetic fields to stimulate nerve cells in the brain.
Why it matters
This FDA clearance introduces a new level of precision and automation to a growing area of psychiatric treatment. By robotically guiding the TMS coil and automatically compensating for patient movement, the Zeta system can improve the consistency and accuracy of the treatment, which is critical for targeting specific brain regions. This could enhance patient outcomes, streamline clinical workflows, and make the therapy more reliable and easier to administer.
This is a prime example of robotics enhancing an existing medical procedure. The robot isn't performing the therapy itself but is acting as a highly precise tool to ensure the therapy is delivered optimally. This kind of assistive robotics has a clear path to adoption in clinical settings, as it augments the capabilities of medical professionals rather than seeking to replace them.
Reimagine Robotics, a startup founded by former Google DeepMind leaders, has emerged from stealth with an AI platform that allows factory workers to train industrial robots through demonstration and correction, without needing to write code. This 'monkey-see, monkey-do' approach uses vision-based imitation learning and on-device continual training to enable robots to learn new tasks in minutes. The company has secured pre-seed funding and is already deploying its robots in manufacturing and electronics disassembly.
Why it matters
This technology directly addresses a major barrier to robot adoption in high-mix, low-volume manufacturing: the high cost and long lead times of expert programming. By empowering non-specialist workers to retask robots quickly, Reimagine's platform could make automation economically viable for a much broader range of companies, particularly small and medium-sized enterprises. It represents a significant step towards more flexible and user-friendly industrial automation.
This is part of a broader trend toward 'no-code' robotics, aiming to democratize the use of complex automation. The focus on on-device learning also reduces reliance on the cloud, which can be critical for latency-sensitive tasks and data privacy. The startup's impressive pedigree from DeepMind suggests a deep understanding of the underlying AI challenges.
The ETH Zurich team we've been tracking for their magnetically guided microrobots in spinal cord repair has developed a new application: navigating the bloodstream for targeted stroke and tumor treatments. The sub-two-millimeter robots use external magnetic fields to deliver medication directly to blood clots or cancer cells, aiming to minimize the systemic side effects of conventional drugs.
Why it matters
This breakthrough represents a significant step towards realizing the promise of medical microrobotics. Enabling highly targeted drug delivery could revolutionize the treatment of life-threatening conditions like stroke, where speed and precision are critical. By concentrating medication exactly where it's needed, this technology has the potential to dramatically improve patient outcomes and reduce the debilitating side effects associated with powerful systemic treatments.
While the technology is still in early stages, it demonstrates a viable method for navigating the complex and dynamic environment of the human circulatory system. The use of external magnetic fields for guidance is a promising approach that avoids the need for onboard power or complex steering mechanisms, simplifying the design of these microscopic devices. The next steps will involve extensive testing to ensure safety and efficacy in more complex biological models.
Researchers at Tel Aviv University have developed microrobots that can navigate complex 3D environments and transport live bacteria using a hybrid magnetic and electric propulsion system. The team describes the capability as '2.5-dimensional navigation,' which allows the bots to move across multiple surfaces and climb over obstacles, a significant improvement over previous designs limited to flat surfaces.
Why it matters
This work expands the operational capability of microrobots, moving them from simple 2D planes into more realistic, complex micro-environments. The ability to precisely transport biological cargo like bacteria in a lab-on-a-chip system opens up new possibilities for automated biological experiments, diagnostics, and the assembly of microscopic structures. It's a key step toward developing more sophisticated tools for interacting with the microscopic world.
The hybrid propulsion system is the key innovation, using electric fields to lift the robot off a surface and magnetic fields to steer it. This allows for a level of control and mobility that was previously difficult to achieve at this scale. The demonstration of transporting E. coli showcases a practical application for this enhanced navigational capability.
Researchers at the Institute of Science Tokyo have discovered that droplets made of DNA can be programmed to move, change shape, and carry cargo using light. By using DNA as a programmable biomolecule that undergoes phase separation, the team created jellyfish-like structures that can be controlled non-invasively, opening new possibilities for microfluidic systems.
Why it matters
This research redefines DNA's role from a simple carrier of genetic information to a versatile, programmable building material for microscopic machines. The ability to control these 'robots' with light offers a non-invasive way to manipulate fluids and transport cargo at the microscale. This has profound implications for targeted drug delivery, lab-on-a-chip technologies, and the fundamental engineering of soft, adaptive materials.
This work bridges the gap between biology and engineering, using the fundamental properties of biomolecules to create dynamic, functional systems. Unlike traditional rigid microrobots, these DNA-based systems are inherently soft and biocompatible, making them potentially ideal for in-vivo applications.
Researchers at the Gwangju Institute of Science and Technology (GIST) in South Korea have developed a 'tactile-embodied' soft robot that can distinguish between external contact and its own bending movements in real time using a single smart material. By embedding magnetically aligned nanowires into the material, the system can differentiate between forces without needing multiple sensors or complex data processing, mimicking a key function of human skin.
Why it matters
This innovation solves a fundamental challenge in soft robotics: discerning self-produced motion from external interaction. Integrating this capability directly into the robot's material reduces complexity, power consumption, and response delays. It paves the way for more autonomous and adaptable soft robots capable of delicate manipulation, navigating confined spaces, and safely interacting with humans, which is critical for applications in medicine, search-and-rescue, and personal care.
This research moves beyond simply adding more sensors to a robot. It's about building intelligence directly into the materials themselves. This 'material intelligence' approach could lead to robots that are not only more capable but also more resilient and cheaper to produce, as complex sensing and processing functions are embodied in the structure itself.
Venture Capital Fuels a Robotics Startup Boom A wave of significant funding rounds is hitting the robotics sector, with nine-figure investments for companies like Neura Robotics, HappyRobot, FieldAI, and PokeBot, alongside notable seed and Series A rounds for specialized startups. This signals strong investor confidence across the stack, from foundational AI to niche industrial applications.
Open-Source Unlocks Advanced Robot Control Major open-source releases are democratizing access to sophisticated robotics software. Robbyant's LingBot-VLA 2.0 offers a model for cross-morphology training, while Robotis has open-sourced its entire humanoid software stack. These initiatives lower the barrier to entry for developing advanced, adaptable robots.
Hardware for Edge AI Proliferates The push for on-device intelligence is accelerating with new hardware releases. NVIDIA launched its more accessible Jetson Thor modules and the Cosmos 3 Edge model for local inference. This trend is reinforced by the expansion of MediaTek's Genio platform and a surge in funding for edge AI chip startups like DeepX.
Microrobotics Demonstrates Tangible Medical Potential Breakthroughs in microrobotics are moving from theory to practical application, particularly in medicine. New developments include light-controlled DNA 'robots' for drug delivery, snail-inspired bots for bowel cancer treatment, and ultrasound-activated devices for remote cell stimulation, all pointing toward a future of highly targeted, minimally invasive therapies.
Soft Robotics Advances with Integrated Sensing and Multifunctionality The field of soft robotics is seeing significant innovation in materials science. Researchers have developed a 'tactile-embodied' material that distinguishes touch from motion, stretchable transistors that can switch functions based on salt concentration, and smart fabrics that mimic human touch, paving the way for more adaptable and human-like robotic systems.
What to Expect
2026-11-16—The 25th IEEE RAS International Conference on Humanoid Robots (Humanoids) begins in Nagoya, Japan.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
626
📖
Read in full
Every article opened, read, and evaluated
250
⭐
Published today
Ranked by importance and verified across sources
20
— 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