Today's Highlights
· China's mandatory autonomous driving national standard lands — L3 systems must include driver takeover-capability monitoring, effective July 2027
· RoboArena score-gaming exposed: a single account handled 58% of Spirit v1.6's evaluations
· Single-leg standing test, 90 moves — eight top-tier general humanoid policies pass zero
1. Research Progress
Eight top-tier general humanoid policies, zero out of 90 single-leg standing moves · locomotion
Humanoid policies have made fast progress on dancing, rolling, and whole-body tracking over the past two years, yet collectively fail at one of the most basic demands: standing still on one leg. On the authors' single-leg balance benchmark, eight published SOTA general policies clean-completed 0 of 90 test moves — they recover from imbalance via stepping and small hops rather than avoiding imbalance in the first place. Achieving "prevention" requires the capture point (a center-of-mass position extrapolated using velocity), which had never driven a real-robot policy before because it depends on base linear velocity that no onboard sensor can provide; this paper finds that expressing it in the support-foot frame cancels this exact term, making it reconstructable from encoders and an IMU alone. Feeding this deployable dynamic center-of-mass observation directly into an actor running on hardware, paired with a reward library translated item-by-item from human posture control, the policy FDDC cleanly completes 86 of 90 held-out moves and transfers to a real Unitree G1. Ablation shows this observation is the single largest contributor — removing it drops performance by 40 points.
Yikai Zhou et al. · arXiv 2608.00500 source
Why does action chunking work? Three popular explanations debunked · manipulation
Predicting and executing multi-step action chunks in a single pass is now nearly standard in robot policies, but why it works has never been settled. After systematically testing in simulation and on real hardware, this paper argues that none of the three mainstream hypotheses — temporal consistency, shortened decision horizon, and representation learning — explains its success; what actually matters is stronger non-Markovian expressiveness and lower compounding error, both of which can, in many settings, be fully reproduced by a "delayed policy" (predicting a single action at each step based on observations from k steps earlier). The authors further identify a previously unnamed benefit — implicit ensembling: an action-chunk policy simultaneously learns multiple temporal relationships such as a_t|o_t and a_t|o_{t-1}, behaviorally equivalent to a model ensemble, which improves robustness and generalization. Building on this, they use a "randomly delayed policy ensemble" to match chunking's performance without chunking at all, and offer a class of policies that explicitly instantiate the ensemble, clearly outperforming chunking itself across multiple domains.
Filippo Lazzati, Kyle Stachowicz, William Chen et al. · arXiv 2608.02547 source
Why VLAs fail on contact-rich tasks: two failure modes, two targeted fixes · vla
Prior remedies mostly added force-sensing branches or training-time regularization, with little investigation into root causes. The authors split failures into two categories: precision failures stem from training misalignment in flow-matching policies, while force failures stem from the structural properties of force signals themselves. Each gets a targeted mechanism, combined into FACT, which achieves an average 66% success rate across five contact-rich tasks versus 41% for the best prior baseline, evaluated over nearly 2,500 real-robot rollouts.
Carlota Parés-Morlans et al. (Stanford / KTH) · arXiv 2608.01402 source
Ego2Robot: converting first-person human video into robot training data at scale — 18,500 hours · vla
Retargeting first-person human video into robot data had previously only been validated for small-scale, single-task policies, and whether it could offer pretraining benefits for VLAs remained an open question. Ego2Robot chains action retargeting, robot-arm visual synthesis, and multi-stage quality filtering into a scalable pipeline that consumes both curated datasets and in-the-wild video, producing 18,561 hours of training data spanning 15 robot embodiments — currently the largest ego-to-robot dataset. To test generalization, the authors add four decoupled perturbation axes on RoboTwin2.0 — visual appearance, scene layout, embodiment, and task semantics — showing that joint pretraining on synthetic and real-robot data consistently improves out-of-distribution generalization across multiple perturbation types, verified in real-robot deployment.
Ye Wang et al. · arXiv 2608.02580 source
Showing ground-truth trajectories to the teacher when labeling driving-VLM chain-of-thought teaches the model to confabulate · autonomy
Today's only paper with community buzz signal (HF↑25). The authors point out that existing autonomous-driving CoT annotation pipelines commonly feed recorded ground-truth future trajectories to the teacher model, inducing a "trajectory anchoring bias" — the teacher doesn't infer decisions from scene evidence but instead fabricates justifications for a known outcome, degrading causal faithfulness and notably worsening hallucination in causally complex scenarios. But simply removing the ground-truth trajectory entangles high-level decisions with geometric synthesis and low-level dynamics. So the authors reframe planning as selecting among explicit candidate trajectories (AD-MCQ) and propose DEFT-RLVR, which turns the future trajectory from an "anchor before the decision" into a "verification target after the decision," improving driving reasoning while preserving or even enhancing general vision capability.
Zixuan Huang et al. · arXiv 2608.01755 source
MiniWorld: a streaming video world model trained from scratch on a single 8-GPU machine in days · world-model
Current video world models mostly repurpose pretrained video generation models, resulting in complex pipelines, expensive compute, and a mismatch between bidirectional pretraining and causal streaming inference; the community has lacked a lightweight, transparent, fully reproducible from-scratch baseline. MiniWorld trains a block-causal Video Diffusion Transformer with flow matching in the latent space of a pretrained Video VAE, building on Diffusion Forcing but with a block-wise non-decreasing noise schedule and two-stage continued training; at inference, it combines a rolling KV cache with pipelined asynchronous denoising. The entire model can be trained in a few days on a single 8-GPU server, with training and inference code plus pretrained weights released together.
Yian Zhao et al. · arXiv 2608.01127 source
3D-printed electrical impedance tomography e-skin aims to free humanoid whole-body touch from "rebuild for every robot" · perception
Traditional taxel arrays are hard to scale up in area, wiring complexity, and robot-specific curvature, which is why humanoid whole-body touch has been slow to roll out. The authors use a geometry-adaptive additive manufacturing process to produce conformal electrical impedance tomography tactile skin: a flexible conductive TPU layer forms a continuous sensing field, contact-induced changes couple with conductive patches to alter boundary voltage, which is then reconstructed with a one-step Gauss-Newton EIT solver. Validated on three prototypes — a flat surface, a U-shaped curved surface, and an iCub facial shape — the curved sensor achieved a mean localization error of 6 mm across 18 contact positions, with no supervised post-processing required.
Haofeng Chen et al. · arXiv 2608.02080 source
Letting machines "suffer a bit internally" first makes them more resilient in unfamiliar environments · locomotion
Biological organisms are composed of diverse sensing-motor parts spanning multiple scales and adapt to new environments extremely fast, whereas machines are built from inert materials at smaller scales and break down when faced with the unexpected. The authors propose and validate a mechanism that hadn't been clearly articulated before: tethering morphologically diverse, originally independent agents together with physical connectors. As the connectors learn to recover behaviors disrupted by their own presence, they trigger and "tame" a sufficiently diverse range of internal perturbations; disruptions later encountered from a new environment happen to fall within this already-tamed range, so the collective can continue normal behavior without any additional learning or adaptation. The more participating individuals and the greater their diversity, the stronger the resilience to new environments.
Alice Hein, Josh Bongard · arXiv 2608.02395 source
Other papers today: Faster-WAM questions whether world-action models truly need deep action modules (arXiv 2608.02365 source); SelfWAM conditions future prediction on executed actions rather than just task prompts (arXiv 2608.00725 source); CoWAM encodes synchronization, role compatibility, and collision convergence as "coordination contracts," intervening in bimanual policies only when necessary (arXiv 2608.02578 source); DynamicWAM uses dual-path motion conditioning for dynamic object manipulation (arXiv 2608.00793 source); an empirical study on asynchronous deployment overlaps model inference with action execution to eliminate stalling (arXiv 2608.01880 source); Open-DiffLoco is the first open-source framework to train deployable quadruped locomotion policies via differentiable simulation, training in 20–60 minutes on a single RTX 5080 (arXiv 2608.02069 source); a survey organizes robot learning along "frozen weights vs. self-written skill code" lines, reviewing 77 representative systems (arXiv 2608.01851 source); the first systematic study of prompt injection attacks on multi-agent robot systems, showing attacks can propagate between agents via shared prompt structures (arXiv 2608.00747 source); Rapid Embodiment Adaptation infers embodiment parameters online from short interaction histories to handle hardware variation (arXiv 2608.01506 source); a language-conditioned latent-space safety filter lets safety constraints be specified in natural language (arXiv 2608.00315 source).
Open Source · Tools · Benchmarks
· Amap (Alibaba's mapping unit) ABot-World-0: Alibaba-owned Amap's general-purpose interactive world model has been upgraded and open-sourced, claiming stable continuous inference for 24 hours on a single consumer-grade GPU with no noticeable degradation in image quality, physical consistency, or interaction responsiveness. Mainstream world models generally cap out at 30 seconds to 1 minute of continuous single-pass generation, with the bottleneck being error accumulation from autoregressive generation; Amap's approach is a LongForcing long-horizon training paradigm — continuously feeding the model's own generated output back as subsequent input during training, while constraining it to stay within a plausible world distribution. ⚠️ Vendor claim source
· Open-source robotic guide dog: a research team has released an open-source design for a robotic guide dog aimed at blind and visually impaired people, pushing this type of assistive form factor from closed products toward a reproducible design source
2. Funding & Deals
Kaiwang Data (KWDATA) | New strategic funding round | Over RMB 100 million · adjacent
Led jointly by the Beijing Yizhuang Industrial Upgrading Fund, Huafang Capital, and Tianji Capital, with Xinding Capital, Mairui Asset Management, and other institutions following on. More notable is who joined the follow-on round: Deep Robotics (Chinese quadruped/humanoid robot maker), Songyan Dynamics (Chinese humanoid startup), AgiBot's spin-off X² Robot, and Zhiyuan Robotics's (Chinese humanoid startup) subsidiary Mifeng Technology — marking the first time this full-chain AI data infrastructure company, founded in 2022, has brought in embodied-AI peers as strategic investors. The company's core business began in autonomous driving data, covering passenger vehicles, commercial vehicles, logistics vehicles, and algorithm firms across the board; over the past two years it has moved into robot data collection, with data categories now covering first-person bare-hand video, tactile sensing gloves, UMI grippers, and third-person panoramic video, deployed in settings including supermarkets, restaurants, homes, and auto production lines. The company says its stable monthly effective-data output capacity is now approaching 100,000 hours, and this round will fund an integrated data trading platform, specialized talent recruitment for world models, and capacity expansion in three key areas. Founder and CEO Yu Xu previously worked at Uber, Momenta, and ByteDance.
Sources: Zhidx source, Cyzone source
Taoshi (Shenzhen) | Series A | RMB 140 million | Valuation surpasses RMB 1 billion · hardware
Participants included Qianhai Financial Holdings, Guangzhou Guokong, Haichuan Juyi, and Hangzhou Zhongshen, with TPG (德太资本) serving as financial advisor. Founded in 2016, Taoshi pursues an unusual niche — micro-scale toroidal-envelope worm-gear reducers — integrating deceleration and 90-degree turning into a single structure, roughly 40% smaller than the conventional "reducer plus separate angle converter" approach, with precision around ±0.5 arc-minutes and a lifespan of up to 10,000 hours. Its products have entered supply chains at Foxconn, Lens Technology, Luxshare Precision, and Piotech, and its dexterous-hand joint modules have already signed a 100,000-unit supply agreement, with customers including Linker Hand and Chaoweit Dynamics. On capacity, its existing ~20,000-square-meter facility produces 500,000 to 700,000 key modules per year, with a target of 1 to 1.5 million once new facilities come online. Actuator systems account for roughly 45% of the cost of a full humanoid robot, making reducers a key link in China-made substitution for imports.
Source: Ifeng Tech · Hardcore source
DISCOVER Robotics | Angel+ round | $100 million · embodied
Participants included IDG, Xinglian, Wuyuefeng, Fortune Venture Capital, Joyoung, Huaying Capital, and the Binhu District Industrial Group, with existing shareholders all adding on and Gengxin Capital serving as lead financial advisor. The notable part is the timing: less than a month after its first angel round of over $100 million closed, the two rounds together total over $200 million; the company positions itself in consumer-grade embodied AI. ⚠️ Single-party claim
Source: Sina Finance source
Unitree Robotics | STAR Market IPO | Underwriter valuation exceeds RMB 50 billion · humanoid
Following the previously disclosed listing valuation of roughly RMB 42 billion, an August 5 book-building date, and an August 10 subscription date, underwriter CITIC Securities has now put out a research-desk assessment putting post-listing valuation at over RMB 50 billion — higher than the valuation implied by the offering terms. The company plans to issue 40.45 million shares, about 10% of post-offering equity, raising RMB 4.202 billion. ⚠️ Broker claim
Sources: Reuters source, Lianhe Zaobao source
Exclaim Robotics (Zurich, Switzerland) | Seed round | €4.29 million · adjacent
The company has come out of stealth with a focus on robots for maintaining AI data centers — in a field crowded with humanoid and factory-automation plays, this is a niche demand pulled directly out of the compute buildout.
Sources: EU-Startups source, AI Insider source
AI² Robotics (Zhipingfang) | Planning Hong Kong listing | Post-money valuation over RMB 20 billion · embodied
Reports suggest the company could IPO as early as next year, with its valuation surpassing RMB 20 billion after its most recent funding round. If it goes through, it would be another embodied-AI foundation-model company, alongside AgiBot and Unitree, putting a valuation payoff on the calendar. ⚠️ Rumor-sourced
Sources: iyiou.com source, The Standard source
3. Commercialization & Deployment
Pony.ai reveals Robotruck production timeline for the first time in two years: 500–1,000 heavy trucks, 100,000 light trucks · autonomy
Over the past two years, Pony.ai's public narrative has been almost entirely dominated by Robotaxi, with its trucking business rarely surfacing. At an August 3 media briefing, the company systematically disclosed its trucking progress for the first time, setting a production target of 500 to 1,000 heavy trucks and 100,000 light trucks. The reason for speaking up now is that cost and engineering maturity have both arrived at once: the seventh-generation Robotaxi fleet has surpassed 1,000 vehicles, and the cost of the self-driving kit has fallen 60%–70%, while the sensor and compute platforms for heavy and light trucks are nearly identical to those used in Robotaxi, letting the truck line directly capture that cost dividend and deployment experience. He Xing, VP and head of the trucking business unit, said the company has accumulated a full set of experience partnering on mass production with OEMs, and that heavy trucks have "reached the stage where large-scale pilot production is possible," with the fourth-generation heavy-truck production line launching alongside a parallel rollout of light trucks. Looking back also explains why it took so long: the first-generation vehicles were "workshop-style" retrofits, and the two prototype vehicles exposed consistency and reliability issues within less than a year of use — "every vehicle had its own unique hardware faults." It wasn't until choosing the FAW Jiefang J7 for the second generation in late 2020 that the biggest pain point of the R&D phase was resolved. ⚠️ Plan-stage claim
Source: 36Kr source
FedEx moves autonomous loading and unloading from pilot to hub scale, Dexterity's dual-arm robots take over trailer loading · industrial
FedEx's partnership with intelligent robotics company Dexterity has scaled up from pilot to full operational deployment: the Mech trailer loading/unloading system has moved into a 325,000-square-foot hub warehouse in Hagerstown, Maryland, at what FedEx describes as a "significantly larger" operating scale than before. The two companies have partnered since 2023, with the predecessor system DexR first deployed at an initial test site in Tracy, California. The dual-arm Mech has a wingspan of nearly 18 feet, a vertical reach of 7.5 feet, and can handle a single payload of up to 60 kilograms, while remaining compact enough to autonomously drive into a trailer to operate. It's powered by Dexterity's physical AI model, Foresight, trained on more than 100 million autonomous actions according to the company, reasoning across three spatial dimensions plus time to balance space utilization, stacking stability, and speed. FedEx says the partnership is testing how physical AI can be woven into broader hub operations — destination planning, trailer allocation, maintenance, and labor workflows.
Source: WWD · Sourcing Journal source
South Korea's a2z lands 11 billion won UAE order, fully replacing Chinese-made autonomous vehicles · autonomy
Autonomous driving company Autonomous a2z has signed a "Smart Mobility Convergence Project" contract with Abu Dhabi's AI company Space42, worth 27.91 million dirhams, roughly 11 billion won. The key detail is the direction of the swap: Space42's TXAI service, previously operated with Chinese-made autonomous vehicles, will be fully replaced with a2z vehicles. The order covers 19 vehicles total, including 8 units of a2z's self-developed L4 driverless shuttle ROii, 5 modified Kia PV5 units, 5 modified Carnival units, and 1 modified MAN bus; the ROii will be manufactured in South Korea and shipped as complete vehicles in the second half of the year, with the rest locally procured and retrofitted. a2z will also build an integrated control facility including a control system and remote driving cabin, and integrate with Space42's ride-hailing app. Shuttle trial operations will begin in the second half of the year on Saadiyat Island and Yas Island, transitioning after 2027 to demand-responsive shuttles and tourist routes. The export had previously been delayed after the China-U.S. conflict with Iran temporarily cut off Middle East sea and air routes.
Source: Seoul Economic Daily source
XPeng's IRON humanoid begins small-batch trial production at Guangzhou plant, mass-production line in final debugging · humanoid
XPeng's humanoid robot IRON has begun small-batch trial production at its Guangzhou plant, with the mass-production assembly line simultaneously entering its final debugging phase; the company plans to reach full mass production in Q4 2026. ⚠️ Vendor claim
Source: Autohome source
South Korean nuclear waste disposal facility deploys seven-robot fleet to keep workers out of high-radiation zones · industrial
A radioactive waste disposal facility in South Korea has deployed a fleet of seven robots to reduce worker radiation exposure — one of the most straightforward cases of robot value in scenarios like nuclear waste handling, where "humans shouldn't linger."
Source: Tech Times source
4. Industry Developments
China releases its first mandatory national standard for advanced autonomous driving: L3 must include driver takeover-capability monitoring, effective July 2027 · autonomy
The "Intelligent Connected Vehicles — Safety Requirements for Automated Driving Systems" (GB 44721—2026), drafted under the organization of China's Ministry of Industry and Information Technology, has been approved and released by the State Administration for Market Regulation and the Standardization Administration of China on July 30, and is set to take effect on July 1, 2027. The standard applies to M-class and N-class vehicles equipped with L3 or L4 systems, but not to automated parking systems. Of its four categories of requirements, the two strictest are: first, the safety level of an automated driving system must at least match that of a qualified, attentive human driver currently performing the dynamic driving task, with specified rules for triggering and executing minimal-risk maneuvers; second, on human-machine interaction, it adds a hard requirement for driver takeover-capability monitoring in L3 systems, requiring the entire activation-and-exit process to be safe and status prompts to be standardized, with automakers also required to clearly communicate capability boundaries and responsibility allocation through official websites, in-vehicle terminals, and other channels. The standard also requires automakers to build safety-assurance capabilities spanning design, development, and post-deployment across four dimensions — safety policy, risk management, safety assurance, and safety improvement — and to build a testing system combining "enterprise assurance-capability inspection + safety-file review + confirmatory testing," with third-party bodies conducting confirmatory assessment across proving-ground, road, and simulation tests. Compared with the UN Global Technical Regulation on Automated Driving Systems (ADS GTR), which China led and which was approved in June 2026, the national standard sets more detailed technical requirements for L3/L4 and builds a unified, standardized test-scenario system. Multiple people interviewed converged on the same point: as regulation shifts from "soft guidance" to "hard constraints," room for exaggerated marketing claims narrows, companies that fail to meet the bar will be weeded out, the supply chain may face a fresh reshuffling, industry competition will shift from a performance race to a safety race, and rising compliance costs will accelerate differentiation.
Sources: 36Kr · NBD Auto source, Beijing Business Today source
RoboArena score-gaming controversy gets a data backing: a single account handled 58% of Spirit v1.6's evaluations · embodied
Following last week's scrutiny of embodied-AI leaderboard credibility, this storyline now has concrete evidence. According to the South China Morning Post, Hangzhou-based Spirit AI (Qianxun Intelligence) briefly topped Nvidia to reach the top of RoboArena in early June with Spirit v1.6 — a benchmark co-built by Nvidia with Stanford, UC Berkeley, and other institutions, evaluating how well general robot policies translate digital instructions into real-world actions. Days later, the leaderboard operators restructured their methodology and removed that model along with several others from the official rankings; also removed was X Square Robot's model, previously ranked fourth. Pranav Atreya, a Berkeley PhD student and lead author on the project, said on X that the team had "retroactively removed evaluation results from organizations found to be engaging in leaderboard manipulation," without naming specific companies. The Chinese tech blog "肉身算法" (Embodied Algorithm), analyzing over 4,600 evaluation records prior to June 3, found that a rater account named ECUST Robot Lab completed 58% of all evaluations for Spirit v1.6, under which the model's win rate was 99%, versus only 66% among independent raters; the blog also claimed that the top five models (including Nvidia's DreamZero) each had at least half of their evaluations coming from a single specific account. The vulnerability lay in the original design: anyone could register as a rater, and evaluated models were hosted on developers' own servers, letting developers selectively power on their servers only when their own team was evaluating, boosting the chance of being matched against their own model. As of Tuesday, Nvidia's DreamZero had reclaimed the top spot, with the remaining top-nine slots held by U.S. players including Google DeepMind and Physical Intelligence.
Source: South China Morning Post source
WeRide sets up internal "Superfluid Lab," reportedly led by a former core DeepSeek multimodal researcher · autonomy
WeRide's newly established Superfluid Lab has begun hiring, with a focus on large-model algorithms, simulation algorithms, and AI infrastructure. According to people familiar with the matter, the lab was formed internally in May this year and has already completed groundwork such as infrastructure restructuring; it is reportedly to be led by Ruan Chong, a former core DeepSeek member and core multimodal-technology researcher — Ruan joined DeepSeek back in the High-Flyer (DeepSeek's parent quant fund) era and was a core contributor to multimodal work including Janus-Pro. The lab centers on foundation-model R&D, with a focus on VLA, while also advancing world-model and multimodal directions in parallel; organizationally it sits independent of the R&D department, without overemphasizing rank or role boundaries, with researchers collaborating around a shared model and jointly accountable for its final capabilities. This reflects WeRide repositioning itself: internally, the company now positions itself as a foundation-model company rather than a driver-assistance supplier. ⚠️ Source: people familiar with the matter
Source: LatePost·Lanjing (Blue Whale Finance) source
Waymo co-CEO: Physical AI has no undo button, "move fast and break things" is over · autonomy
In an interview at Y Combinator, Waymo co-CEO Dmitri Dolgov said that once AI moves from digital products into machines operating around people, Silicon Valley's familiar playbook is no longer acceptable and should be replaced with "move fast, ship safely" — safety must be built into the model, training methods, and system architecture from the start, not bolted on after deployment. His framing: a chatbot error can usually just be retried, but vehicle failures cost "lives, not tokens", with "no undo and retry button at all." The weight behind this comes from Waymo's own scale and scar tissue: the company has logged over 220 million fully driverless miles and delivers over 500,000 fully autonomous rides per week; a study comparing 56.7 million fully driverless miles against a human baseline showed statistically lower rates of injury reports and airbag deployments. But in May the company recalled nearly 3,800 vehicles because its software could potentially direct vehicles into flooded roads, and later recalled nearly 3,900 vehicles and temporarily paused highway service after more than a dozen incidents of vehicles entering closed highway construction zones.
Source: Benzinga source
Xu Xin: without spending $1 billion a year, you don't get a ticket to embodied AI · adjacent
In an interview, Xu Xin (founder of Capital Today, a prominent Chinese VC), said embodied AI is harder than robot hardware itself, because the "brain" component demands extremely heavy compute and data — not every team can pull it off. She put the threshold at $1 billion a year. ⚠️ Interview statement
Source: CNFOL source
The spillover effects of U.S. robot import restrictions: South Korean parts suppliers named as beneficiaries · hardware
More than a week after the FCC's covered-list expansion and import restrictions took effect, Korean media are now tallying a different set of consequences: as the U.S. reshoring policy squeezes out Chinese suppliers, it's opening space for Korean parts makers, with KB Securities naming LG Innotek as a beneficiary with improved prospects. China's Ministry of Commerce has previously said it would firmly retaliate, and neither side's posture has shifted this week. ⚠️ Broker claim
Source: Chosunbiz sourcesource
Hardware · Supply Chain
· Samsung Electro-Mechanics, LG Innotek: the two companies have jointly secured exclusive supply of camera modules for Tesla's Cybercab — Robotaxi mass production is turning exclusive sensing-hardware slots into firm orders source
· igus energy chains: launched a robot energy chain supporting 600-degree rotation, targeting wear and interference in industrial robot wrist and turntable cabling under large-angle rotation source
· Grinm Advanced Materials (Youyan): net profit up over 40% year-on-year in H1, with the company saying NdFeB magnet demand from humanoid robots could become a new growth driver for rare-earth magnetic materials — for now this remains a forecast rather than actual orders ⚠️ Forward-looking claim source
Top comments (0)