Author | Cen Feng, Zhang Xian Editor | Cen Feng
IROS 2026 has returned to Pittsburgh.
Known as "Roboburgh," this city is home to Carnegie Mellon University (CMU) and a holy land for global autonomous driving and robotics research. It has been 31 years since IROS was last held here. Returning to its old haunt after three decades, the conference has undergone a fundamental shift in temperament: the trend has completely shifted from "making robots smarter (stronger models)" to "making robots more capable (stronger physical entities)."
Over the past two years, the main narrative of embodied AI has been "large models entering robots," attempting to stack up intelligence through massive amounts of text and video tokens. However, at this year's IROS, various signs indicate: it is far from enough for models to understand semantics; robots must understand physics. This means the focus of industrial competition is undergoing a massive migration from "algorithmic imagination" to "physical execution capability." It is precisely at this turning point that Chinese companies, leveraging the dual advantages of manufacturing dividends and deep scenario applications, have demonstrated a "saturated" presence at the IROS venue.
01. The Metaphor of Pittsburgh From "Steel City" to "Physical World Simulator"
The transformation of Pittsburgh itself is a grand metaphor. This former "Steel City" reinvented itself through the research spillover from CMU after the decline of its manufacturing industry. Its underlying tone is not the light and agile internet, but heavy industry, precision manufacturing, and extreme engineering in real-world scenarios.
By setting the 2026 venue here, IROS actually sent a signal: the next leg of robotics is not in the cloud, but in muddy workshops, complex warehouses, and highly dynamic hospitals. This is highly aligned with the shift in the conference's technical agenda. The highly anticipated workshop, "Physical World Models for Scaling Embodied AI," led by top scholars such as Jiajun Wu from Stanford, no longer asks "Can models generate realistic videos?" but rather "Can models embody real physical consistency?"
The concurrently held WorldArena 2.0 Challenge integrated visuo-tactile perception and real physical manipulation into a unified evaluation for the first time. The evaluation criteria have undergone a qualitative change: "semantic correctness" gives way to "physical correctness."
To summarize in one sentence: embodied AI has hit a "data wall," and the tool to climb over it is no longer just computing power, but physical entities. Whoever can produce physical interaction data at scale holds the hard currency leading to AGI. And this is exactly within the "firing range" of China's robotics industry.
02. Humanoid Embodiments From "Solo Overseas Expansion" to "Platform-based Siege"
Humanoid robots are the most prominent Chinese symbol in this year's IROS exhibition hall. As a Gold Sponsor, Galbot showcased its robot Galbot ET1, its first humanoid robot specifically designed for bipedal locomotion with a complete bipedal form. At the IROS venue, Galbot ET1 demonstrated high-dynamic motion control capabilities through continuous dance performances, attracting many attendees to stop and watch.
Galbot ET1. On the other hand, Zhijian Dynamics' self-developed humanoid robot i7 Pro and all-scenario service robot iX focus on working capabilities in industrial scenarios. At the venue, the real i7 Pro machine completed tasks such as sealing cartons and packing physical objects. Relying on the SimpleWAM world action model deployed on the edge, it fully demonstrated the robot's practical working effects in real production line tasks.
Zhijian Dynamics i7 Pro. Open-source robotics enterprise RoboParty brought the brand-new RP1 full-stack open-source humanoid robot to the IROS venue. Built for developers and model researchers, RP1 features stronger motion capabilities and a higher degree of modularity, adopting an open architecture suitable for scientific research exploration and secondary development. According to the staff on site, RoboParty will continue to open up the full-stack capabilities of RP1, covering hardware drawings, control code, engineering documentation, and the Sim2Real development process.
In addition, in the quadruped robot sector, ZhiShen Technology, Deep Robotics, and Dobot also made appearances at this IROS. Among them, ZhiShen Technology exhibited the product matrix of the ZhiShen Gangbeng L1 series: ZhiShen Gangbeng L1, ZhiShen Gangbeng L1-W, the new L1 Maker for educational training, the new-generation ZhiShen Gangbeng L2 series, and 8 self-developed joint modules.
The bigger picture is that Chinese robotics companies are forming an "army group" effect. In addition to the embodiment companies, a group of domestic robotics emerging forces, including star enterprises such as Star Era, YuanSheng Intelligence, BrainCo, LingChu Intelligence, RoboSense, JianZhi Robot, Inspire Robots, Noitom, and LightWheel AI, scattered throughout the venue with their core components, data platforms, and data collection equipment solutions, gaining high attention. Chinese power is completing an "all-round coverage of ecological niches" in the humanoid robot track, which was once dominated by Boston Dynamics and Honda.
03. Dexterous Hands The "Chinese Matrix" Reshaping Tactile Feel
If the humanoid embodiment is the most visible flag, then the dexterous hand is the densest "Chinese muscle." The collective appearance of a batch of dexterous hand exhibitors, including Star Era, YuanSheng Intelligence, DH-Robotics, BrainCo, LinkerBot, Critical Point, Wuxi LingZhang, and Inspire Robots, marks China's dominance in the dexterous hand track.
YuanSheng Intelligence Apex Hand. The Apex Hand presented by YuanSheng Intelligence this time focused on demonstrating the comprehensive balance of the Apex Hand among "core capabilities such as dexterity, reliability, response speed, load capacity, tactile perception, and control accuracy" through demos such as long-duration heavy object grasping and tool usage. In addition, YuanSheng Intelligence also released the "YuanChuang Plan" targeting global universities, research institutions, and developer teams, offering long-term free prototype trials, a full set of development toolkits, and dedicated technical support.
XHAND 1 PRO. Star Era brought the new dexterous hand XHAND 1 PRO, equipped with 21 active degrees of freedom (DOF) across the whole machine. It adopts a fully joint back-drivable design, allowing buffering and yielding under external forces for smoother and safer operation; an independent DOF is added to the side palm to enhance the thumb-pinky opposition ability, enabling the reproduction of more complex human-like hand movements. Another dexterous hand enterprise, DH-Robotics, brought new products to this IROS, including the ADH-5-13 fully direct-drive five-finger dexterous hand and the UDH-3-7 three-finger dexterous hand, comprehensively covering the complete end-effector matrix of two-finger, three-finger, and five-finger hands.
DH-Robotics ADH-5-13 Dexterous Hand. LingZhang Robot brought multiple dexterous hand products, including the LZ-Pro for high-load scenarios and the 23-DOF high-degree-of-freedom dexterous hand LZ-Explore. On site, the LZ-Pro performed dumbbell grasping and lifting demonstrations. This product adopts a tendon-driven scheme, features 11 degrees of freedom, and has a maximum load capacity of up to 30kg. While ensuring dexterous manipulation, it also targets high-load application scenarios such as industrial handling and special operations. In the dexterous hand track, Chinese companies are "besieging" the global market through different technical routes.
04. Hardware Closed Loop The "Full-Perception" Explosion of the Tactile Matrix
The most surprising change at this year's IROS is: tactile sensing is no longer a vassal to vision, but has become the core of the action closed loop. At the same time, Chinese tactile enterprises are forming a "sea, land, and air" style of coverage. Daimon Robot made its debut with the Daimon-TWM tactile-anchored world model. On-site demonstrations included three flexible manipulation demos: weaving friendship bracelets, heat-transfer canvas bags, and pencil case storage. Relying on tactile perception to anchor physical interaction, it achieves fine manipulation of soft materials such as fabrics, demonstrating the model's physical AI capabilities in life-like dexterous operations.
Bracelet weaving demonstration. TaShan Technology showcased its independently developed multi-series sensors, including the TS-F fingertip tactile sensor and the TS-E manipulator tactile sensor, which can achieve proximity detection and high-frequency tactile feedback, covering the complete perception process from proximity, contact, to grasping. TuJian Technology demonstrated the complete product and technology closed loop of "hardware - data - model" in the tactile track on site. The solution expands from electronic skin to full-stack tactile perception, including iteratively upgraded multimodal flexible electronic skin, crumple-able planar devices, as well as the TachinHand full-hand tactile perception solution and TachinGlove tactile data collection hardware.
Leveraging its status as a Platinum Sponsor for this IROS, Paxini showcased the mass production capabilities of its multi-dimensional force tactile sensor arrays on a large scale to the industry, academia, and research sectors. ShiYue Technology also brought high-resolution flexible electronic skin. The collective explosion of this "tactile matrix" marks that Chinese enterprises have filled the last gap in physical feedback for embodied AI.
05. Data Infrastructure The Chinese Version of "Data Alchemy"
If the dexterous hand solves "whether the hand can do the work," then the data infrastructure solves "how the hand learns to do the work." At this IROS, Chinese data enterprises have elevated the word "infrastructure" to a new magnitude. Skyworth XR brought a full set of first-person perspective data collection hardware, among which the E2 is a mass-produced lightweight model weighing less than 140g, equipped with dual 1080P RGB cameras, suitable for university scientific research. The main mass-produced model E6 adopts the Qualcomm XR2 Gen1 platform, equipped with 4 grayscale global shutter cameras + 2 RGB cameras, supporting large-scale robot data collection.
Skyworth XR. In terms of data providers, LightWheel AI focuses on the data end, showcasing its released EgoSuite-Open100K, the world's first 100,000-hour-level full-modality human behavior open-source dataset. This dataset covers 128 categories of scenarios and over 15,000 tasks, and is simultaneously open-sourced to lower the threshold for acquiring embodied AI data. In addition to data collection hardware, JianZhi Robot also showcased the construction of the Human Data multimodal data foundation, bringing the collaborative operation mode of Hardware, DFM, and Data to the venue, and presenting how human operational experience is transformed into trainable resources for models through the Whole-Body Mesh and Gen-HumanEgo datasets.
Boden Intelligence showcased its data infrastructure capabilities in full-stack Physical AI and its integrated hardware and software achievements. From collection equipment to data factories, and then to compliant circulation, the entire data chain relies on China's massive industrial advantages, and the cluster advantages of domestic enterprises are almost impossible to replicate in the short term.
06. Academic Methodology From "Followers" to "Definers"
The breakthroughs of Chinese power at IROS 2026 are penetrating into the underlying genes of algorithms. The paper "Learning Athletic Humanoid Tennis Skills from Imperfect Human Motion Data" by the Galbot team won the IROS Best Entertainment and Amusement Paper Award. Tennis is a high-difficulty whole-body motion task for humanoid robots. This research extracts and combines motion skills from flawed human motion data, enabling the robot to perform ball tracking, movement, center of gravity adjustment, and timing control, achieving whole-body coordinated hitting.
The "Eye of Motion" paradigm proposed by the CMG Lion Rock AI Lab directly hits the fatal flaw of VLA models: "shortcut learning" caused by fixed perspectives. Through dynamic perspective reconstruction, they improved the success rate from 43% to 83% without altering the model backbone. This approach of solving "algorithmic generalization" at the "data layer" perfectly fits the technical undertone of IROS 2026.
In addition, the MIWA framework from the Dalian University of Technology HPC team drew inspiration from marionettes, utilizing a progressive annealing mechanism to solve the "conservative strategy" problem in high-dynamic motion, successfully replicating complex dances from online videos on the Unitree G1. These works prove that Chinese scholars have begun to use Chinese hardware platforms to define general control problems in global robotics.
07. Conclusion Physical Capability, the Ticket to the Second Half of AGI
Returning to Pittsburgh after 31 years, IROS 2026 sent a clear signal: the competition in embodied AI has shifted from an "IQ contest" to a "physical endurance game."
Large language models can acquire intelligence through "scaled reading," but robots must acquire a soul through "scaled collision" and "scaled interaction." These experiences are not in the texts of the internet, but in every friction and collision in factories, kitchens, and laboratories. The confidence of Chinese companies in this game does not lie in how elegant the code they write is, but in the fact that we have places to produce data, scenarios to verify algorithms, and industrial chains to turn panacea-like intelligence into democratized costs. The "steel" trait of Pittsburgh reminds us: the root of robotics is forever in manufacturing and engineering. And when this root is rediscovered by the whole world, the position where Chinese companies stand is closer to the eye of the storm than ever before.