Today's Highlights
· Unitree Robotics' STAR Market IPO registration was approved (China Securities Regulatory Commission consent), planning to raise RMB 4.202 billion; at roughly 104 days from acceptance to registration effectiveness, this is the fastest pre-review on record for the STAR Market, putting China's "first listed humanoid robot company" on the horizon.
· World models are shifting from "contestant" to "infrastructure": D-Robotics' Uranus achieves frame-level closed-loop control, Zhiyuan Robotics (Chinese humanoid startup) released an open general-purpose training ground, and multiple world-model papers (Valdi, ABot-M0.5, etc.) landed the same day, extending the recent capital-attracting wave.
· Home embodied AI draws another major bet: Lexiang Technology (Chinese home robotics startup founded by former Dreame executives), completed a nearly RMB 500 million Pre-A round (led by Ant Group), bringing total funding to RMB 1 billion; Ant Group has now backed 12 humanoid/embodied AI companies in 18 months.
· Global bearing leader SKF formed a joint venture with Chinese harmonic reducer leader Leaderdrive, with SKF holding a 60% stake — moving directly into precision joint transmission for humanoid robots.
· The first Shanghai International Embodied Intelligence Expo opened; concurrently, the China Embodied Intelligence Industry Development Report (2026) stated China is among the world's fastest-growing embodied AI markets, with an annual compound growth rate of roughly 22%–23%.
I. Research Progress
"Domain arithmetic" adapts VLA to environmental drift with a single sample · vla
VLA models often fail wholesale when a camera pose changes or the robot changes (e.g., Panda→UR5e); traditional adaptation requires collecting multiple demonstrations per task, which is costly. This paper characterizes environmental drift as a "domain vector" that can be added or subtracted in weight space, achieving cross-domain adaptation with just a single sample — substantially reducing reliance on target-domain demonstration data. It topped the day's Hugging Face trending list with a score of 16.
Taewook Kang et al. · arXiv 2607.00666 source
Letting world models complete 3D point clouds for more stable long-horizon prediction · world-model
Large video dynamics models lack explicit 3D structure, causing geometric drift and error accumulation over long rollouts. This work injects 3D spatial structure into world models via point completion, outperforming 2D video baselines on both dynamics-learning accuracy and long-horizon consistency — aligning with the day's main theme of "world models moving toward usability."
Skand Peri et al. (Oregon State) · arXiv 2607.00148 source
A structured 4D latent prediction model for robot planning · world-model
Most video prediction models remain confined to 2D sequences, lacking 3D geometry and struggling with precise spatial reasoning. This work (including Yilun Du) builds a structured 4D (3D + time) latent prediction model, providing physically consistent future predictions for robot planning.
Zhiyi Li et al. · arXiv 2607.01166 source
ABot-M0.5: a world-action model unifying mobility and manipulation · world-model
Mobile manipulation remains difficult for current embodied learning approaches. ABot-M0.5 centers on "three-layer alignment" (temporal granularity, action space, and train-test consistency), using intermediate implicit actions to bridge video latents and embodied control — unifying mobility and manipulation within a single world-action model. Hugging Face trending score: 8.
Ronghan Chen et al. · arXiv 2607.00678 source · Analysis: Shuyuan AI source (WeChat, CN)
Valdi: a value-diffusion world model that jointly predicts value, reward, and dynamics · world-model
Diffusion models can capture uncertain future dynamics, but iterative inference is slow and ill-suited to low-latency latent-space planning. Valdi trains a latent diffusion dynamics model end-to-end within an online control loop, TD-MPC style, simultaneously predicting the value function, reward, and dynamics — balancing expressiveness with online usability.
Christopher Lindenberg, Kashyap Chitta · arXiv 2607.00917 source · Analysis: Shuyuan AI source (WeChat, CN)
Letting world models act as "referee": RoboWorld's neural simulation evaluates general-purpose policies · benchmark
Using video world models to evaluate general-purpose robot policies can bypass real-robot deployment, but the world model's own errors make rollouts unreliable, and inference is slow. RoboWorld targets policy evaluation that is both fast and reliable, echoing this week's discussion of "world models not as contestants, but as referees."
Byeongguk Jeon et al. (KAIST) · arXiv 2607.01060 source
Transferring human-hand tactile pretraining to dexterous manipulation · perception
Touch provides force feedback that vision cannot replace for contact-rich tasks, but tactile-equipped datasets are small in scale and narrow in contact coverage. This paper does human-centric, transferable tactile pretraining, alleviating the tactile data bottleneck in dexterous manipulation — resonating with the day's momentum around "tactile foundation models."
Chi Zhang et al. · arXiv 2607.01067 source
Fei-Fei Li, Jim Fan, and Danfei Xu: dexterous-hand data may have "gone down the wrong path" · manipulation
The three authors co-wrote a piece reflecting on the current approach to dexterous-hand learning, and released a 100-hour bimanual dexterous-hand teleoperation dataset covering over 200 everyday objects, 22 motion primitives (grasping, squeezing, insertion, wiping, folding, etc.), and more than 7,700 trajectories — arguing for reconstructing dexterous manipulation using data that better reflects human manipulation priors.
Fei-Fei Li, Jim Fan, Danfei Xu, et al. · Analysis: DeepTech source (WeChat, CN)
FurnitureVLA: a VLA for real-scale bimanual furniture assembly · vla
Existing robotic furniture assembly work mostly remains at toy scale or single-arm. FurnitureVLA is the first systematic study of a VLA for real-scale bimanual furniture assembly, pushing long-horizon two-handed manipulation closer to practical scenarios.
Chenyang Ma et al. · arXiv 2607.01212 source
Other papers today: FAR, enabling test-time recovery and continuous improvement via failure-aware retry (arXiv 2607.01111 source); ROSA, a foundation-model serving system for robot factories (arXiv 2607.01088 source); the paper for NVIDIA GEAR's self-improving skill-discovery system ASPIRE is now up (previously reported, arXiv 2607.00272 source); Wake up for Touch, mask-isolated tactile alignment in MLLMs (arXiv 2607.00302 source); From World Models to World Action Models, a tutorial survey of robot world models (arXiv 2607.00836 source); Path Planning in Physically Viable World Models (arXiv 2607.00673 source).
Open Source · Tools · Benchmarks
· Zhiyuan Robotics' General Training Ground: connects real-robot data, simulation evaluation, world models, and robot deployment into one open platform for community training and evaluation of embodied policies, aligning with the day's several "world model as evaluator" papers source (WeChat, CN)
· AgiBot's X-Tokenizer: a cross-modal embodied action tokenizer; the company claims a 13.5% improvement in multimodal alignment and an 8.25% improvement in long-horizon task performance, alongside a new QUANXTA Zero embodiment-free data collection solution source
II. Funding & Deals
Lexiang Technology (Zeroth) | Pre-A | Nearly RMB 500 million | RMB 1 billion total · embodied
Led by Ant Group. Lexiang was founded by former Dreame executives (Chinese cleaning-robot maker) just a year and a half ago; this round coincided with the launch of its home embodied-AI brand "Zeroth" (spanning collaborative, companion, humanoid, and other product forms). The company claims first-half revenue growth of roughly 600% and about 30,000 orders. Home scenarios are the new focus of the current wave of embodied AI capital.Source: Robotics Outlook source (WeChat, CN)
CarbonSix (South Korea) | Series A | $40 million (about KRW 60 billion) · industrial
Co-led by DSC Investment and LB Investment, with participation from KDB, IMM, and all existing seed-round shareholders. CarbonSix builds physical AI systems for factories (robotic intelligence software plus manipulators/end effectors), using a "data flywheel" approach that collects production-line data during deployment to feed back into the model; the company says it already has commercial contracts and revenue.Source: Chosunbiz source
Agility Robotics | SPAC listing | ~$2.5 billion valuation · humanoid
U.S. humanoid robot maker Agility plans to merge with Churchill Capital XI to go public under ticker AGLT, with the deal bringing in over $600 million in funding (including a PIPE led by Foxconn). Its bipedal robot Digit is already deployed at 9 customer sites including Schaeffler, GXO, and Toyota Canada, with over $300 million in multi-year orders on hand — positioning it as one of the few "pure-play humanoid" names on U.S. markets.Source: MarketBeat source
VAST | New round | Over RMB 1 billion · world-model
VAST, which started out in 3D generation, has raised over RMB 1 billion in a new round and is pivoting toward world models, aiming to use world models to support a low-barrier interactive content platform. Geely Capital invested in two embodied AI/world-model companies in a single day, with this round being one of them.Source: GeekPark source (WeChat, CN)
Kunlunxing Robotics | Three consecutive rounds | Tens of billions of RMB cumulative · embodied
Backed by Casstar (Zhongke Chuangxing), Hillhouse Ventures, Banyan Capital, Zhending Capital, and others. The company says it reached unicorn valuation within 90 days of registration, setting a new record for the fastest-growing embodied AI startup in China. The frequent appearance of such "unicorn-at-founding" narratives is itself a sign of overheating in the current sector.Source: Beijing Fund Industry Association source (WeChat, CN)
Luxonis | Series A | $14 million · hardware
Focused on the perception layer of physical AI (depth/spatial vision hardware and software). Compared with the recent cluster of "robot brain" funding rounds, this round directs capital toward the more upstream, less-covered sensing and perception stack.Source: The Robot Report source
III. Commercialization & Deployment
Pony.ai says its Guangzhou and Shenzhen Robotaxi operations have each reached per-vehicle profitability · autonomy ⚠️ Company claim
Pony.ai reported results claiming its Robotaxi operations in Guangzhou and Shenzhen have successively achieved per-vehicle profitability; the same day, the GAC Aion Hyper HT Robotaxi it backs began deliveries. Per-vehicle profitability is a key milestone for Robotaxi unit economics, but this is currently a company self-report, with methodology and depreciation assumptions not independently verified.Source: Autohome source
Blattner awards Built Robotics a $75 million contract · industrial
According to The Robot Report, energy engineering giant Blattner is deploying Built Robotics' physical AI robots at scale in U.S. solar farm construction under a contract worth roughly $75 million, handling high-risk tasks such as pile driving, surveying, and trenching, with operators supervising from a safe distance — pointing to real order flow driven by "clean-energy demand plus construction cost-and-safety improvements."Source: The Robot Report source
AgiBot's A3 humanoid lands in Europe, launches Robot-as-a-Service in the UK · humanoid
AgiBot debuted its A3 humanoid robot in Europe and launched a Robot-as-a-Service (RaaS) rental model in the UK. Following mass production at the ten-thousand-unit scale in China, leading manufacturers are beginning to treat overseas expansion and "rent-as-needed" as part of their scaling strategy.Source: Robotics & Automation News source
UBTECH humanoid robots deployed at a China border crossing for crowd control · humanoid
According to foreign media, UBTECH humanoid robots have been deployed at a busy international border crossing in China to assist with crowd flow/order management. Compared with home-companion use cases, this type of government/public-facing scenario is one of UBTECH's humanoid deployments that comes closer to real working conditions.Source: Interesting Engineering source
Novares partners with Innov8 to pilot humanoid robots in auto-parts manufacturing · humanoid ⚠️ Pilot-stage claim
Auto-parts maker Novares has partnered with Innov8 to pilot humanoid robots on an auto-parts production line. This is an early pilot, not a scaled deployment, and its real-world viability remains to be seen.Source: Ifeng Tech source
IV. Industry Developments
SKF and Leaderdrive form joint venture, entering precision joints for humanoid robots · hardware
Global bearing and transmission leader SKF and Chinese harmonic reducer leader Leaderdrive signed an agreement to establish a joint venture in China, with SKF holding a 60% stake, focused on high-precision transmission components for robot joints; operations are expected to begin by the end of 2026. The joint venture is anchored in China — the world's largest humanoid robot market — while leveraging SKF's global network to reach Europe, Japan, and the United States, marking a significant step by a foreign giant into humanoid joints, the "most expensive component."Source: PR Newswire source
D-Robotics' Uranus world model achieves frame-level closed-loop control · world-model
D-Robotics (a Horizon Robotics affiliate) released its Uranus world model, moving from "video generation" to "interactive simulation" with frame-level closed-loop capability. Unlike most manufacturers that position world models as an "embodied brain," Uranus is positioned as infrastructure for robot development (data generation, policy evaluation), echoing the day's industry-wide shift toward "world models as referees, not contestants."Source: Sohu source · QbitAI source (WeChat, CN)
Ant Group doubles down on humanoid robots: 12 deals in 18 months · adjacent
According to CNBC, Ant Group has invested in 12 humanoid/embodied AI companies since early 2025, with its lead investment in Lexiang Technology's nearly RMB 500 million Pre-A round the latest example. Capital from internet giants is accelerating into the application and full-robot layers of the embodied AI industry.Source: CNBC source
Beijing Humanoid Robot Innovation Center unveils "world's first full-scale, whole-body VLA technical framework" · humanoid ⚠️ Company claim
The Beijing Humanoid Robot Innovation Center says its "Wisdom Sail" (Huisi Kaiwu) platform has launched a full-scale, whole-body VLA framework, primarily addressing gait oscillation and motion instability issues that arise when mainstream VLA models are transferred to high-dimensional, whole-body humanoid control. The "world's first" claim is the company's own description; its real-world capability and production readiness remain to be verified.Source: Beijing Humanoid Robot Innovation Center source (WeChat, CN)
Musk shares group photo of Tesla Optimus mass-production team · humanoid ⚠️ Statement/claim
Musk posed with the Optimus production-line team at the Fremont factory (formerly producing the Model S/X, which ceased production in May), and acknowledged that humanoid mass production "will start out extremely slow, unlike car manufacturing." This reflects production-line conversion progress and commentary, with no mass-production delivery figures yet given.Source: Humanoid Robot Alliance source (WeChat, CN)
International Federation of Robotics (IFR) elects new president · adjacent
Jane Heffner of Teradyne was elected as the IFR's new president, completing a leadership transition at the industry governance level.Source: AI Insider source
Hardware · Supply Chain
· MEMS six-axis force sensor chip: Zhongke Midian says it has begun mass production of a MEMS six-axis force sensor chip, aiming to substantially cut the cost of six-axis force sensors — currently RMB 80,000–120,000 per unit, with each humanoid robot requiring 4–6 units at roughly RMB 20,000 each source (WeChat, CN)
· Joint-specific chip: China's first dedicated gallium-nitride magnetic encoder chip for humanoid robot joints has been released, claimed to cut power consumption by 22%, filling a domestic gap in high-end robot sensing source (WeChat, CN)
· Weihong Corporation: its micro servo motors have entered the North American dexterous-hand supply chain, becoming a "new favorite" among overseas manufacturers source
· Dexterous-hand valuation paradox: RealMan (Lingxin Qiaoshou), a Chinese dexterous-hand maker, has built a roughly RMB 43 billion valuation on the narrative of holding "over 80% global market share in high-DOF dexterous hands," while Yinshi Robotics, the actual shipment volume leader, has a comparatively modest valuation — capital is paying a premium for narrative, not for production volume source (WeChat, CN)
V. Weekly Watch
China rolls out a wave of humanoid robot standards and field-training policies · hardware
Regulators were highly active on humanoid robots this week: China's Ministry of Industry and Information Technology (MIIT) opened public comment on an industry standard for six-axis force sensors for humanoid robots, bringing this core force-control component into standardization; MIIT and the State-owned Assets Supervision and Administration Commission (SASAC) jointly launched a 2026 special initiative for humanoid robot and embodied AI field-scenario training, specifying that core products should see routine deployment across multiple typical scenarios by year-end; Hangzhou's Binjiang district and other locations also rolled out China's first humanoid robot quality-management standard (with Unitree and others participating in its drafting). The common thread: using standards and application scenarios to push the industry from "usable" to "good to use."Source: Home Service Robot Committee source (WeChat, CN) · Dajishi source (WeChat, CN)
Morgan Stanley: China's humanoid robot fleet could reach 1 billion units by 2050 · humanoid ⚠️ Single-firm long-range forecast
Morgan Stanley's China Humanoid Robot Industry report offers a long-range projection: by 2050, China's humanoid robot fleet could reach roughly 1 billion units, corresponding to a market of about $7.5 trillion; the bank also raised its forecast for China's 2026 humanoid shipment volume. This is a single investment bank's long-term scenario, using a different methodology from the "22%–23% compound annual growth" figure cited the same day in the China Embodied Intelligence Industry Development Report (2026) — best read as a directional trend rather than a firm number.Source: Robotics Knowledge Base source (WeChat, CN)
This week in supply chain: harmonic reducers see "foreign entry + Chinese duopoly" jockeying simultaneously · hardware
Harmonic reducers were the hottest supply-chain segment this week: on one side, a foreign giant is formally entering the market — SKF taking a 60% stake in a joint venture with Leaderdrive; on the other, China's duopoly is solidifying — Leaderdrive holds over 60% of the Chinese market and ranks second globally, while Laifu Harmonic just listed on the Hong Kong Stock Exchange as "the first harmonic reducer stock in Hong Kong" (shipping nearly 292,000 units in 2025). Industry data shows global harmonic reducer sales reaching roughly $850 million in 2026, with China's 2025 market at about RMB 1.6 billion and consumption of roughly 957,500 units — humanoid robot mass production is the main source of incremental demand.Source: Shenlian Industry Research source (WeChat, CN) · Industry Energy Circle source (WeChat, CN)
This week in supply chain: six-axis force sensors and robot chips — cost reduction versus the debate over "whether to build dedicated chips" · hardware
Six-axis force sensors remain the most prominent bottleneck in mass-production costs — each humanoid robot needs 4–6 units, at a total cost of roughly RMB 80,000–120,000 per unit — and this week saw parallel efforts to ease that via "MIIT's industry standard plus MEMS six-axis force chips lowering costs." On the chip side, automakers are collectively crossing over (Xpeng's Turing chip, MetaX's joint venture with UBTECH, ASR Microelectronics' 4nm partnership with D-Robotics) to jockey for position in joint control and compute; but Song Jiqiang, head of Intel China Research Institute, cautioned that now is not the best time to launch dedicated robot chips — VLA capabilities are still evolving, and general-purpose chips adapt better to algorithmic uncertainty — offering a note of caution amid the current "chip-building fever."Source: Bengbu Midian Technology source (WeChat, CN) · Gasgoo Embodied Intelligence source (WeChat, CN)
Top comments (0)