DEV Community

Shawn
Shawn

Posted on

FutureX · Physical AI Daily — Issue 59 (07/16)

Today's Highlights

· Xiaomi open-sources 38-billion-parameter world foundation model U0, topping the World Arena leaderboard for embodied video generation

· Walden exits stealth with $300 million backed by Toyota and Nvidia, $1.1 billion valuation

· RoboParty closes two rounds totaling nearly RMB 500 million, with CATL leading exclusively

· XPeng's second-generation VLA passes German certification without retraining, sprinting toward Europe deployment

· LimX Dynamics releases COSA 0.5 humanoid brain system, argues "the model is not the brain"

I. Research Papers

Xiaomi U0: Turning a world foundation model directly into an embodied synthesis engine · world-model

Following yesterday's preview of "world models as data factories," Xiaomi officially open-sourced the model and released the paper today. U0 is a 38-billion-parameter multimodal autoregressive model that unifies text-to-image generation, image editing, embodied scene generation, cross-embodiment transfer, and embodied video generation into a single framework via unified next-token prediction, initialized from the open-source EMU3.5 (Qwen-3-32B). It is the first to support high-quality multi-view scene generation across multiple robot embodiments, ranks first on the World Arena benchmark for embodied video generation, and lifts the generalist policy π₀.₅'s out-of-distribution success rate on real-world manipulation tasks from 36.9% to 63.2%; paired with anti-diagonal decoding, 1024² image generation is roughly 83x faster than native autoregressive generation. The key takeaway is that it validates the idea that a foundation world model can serve both as an embodied world model and as a scalable data engine.

Xiaomi Robotics · arXiv 2607.11643 source · Analysis: Shuyuan AI source (WeChat, CN)

FlowWAM: Using optical flow as a "universal action language" to unify world action models · world-model

Uses optical flow as a unified action representation for world action models (WAM), aligning well with pretrained video generators while carrying enough motion cues for control—an attempt to answer "how to find an action representation for video generators that is both alignable and drivable."

Yixiang Chen et al. · arXiv 2607.13017 source

VistaVLA: Adding an explicit 3D scene representation to VLA · vla

Existing VLAs map language and 2D images directly to actions, lacking scene-level 3D representations and struggling to reason about spatial layout and geometric constraints. VistaVLA uses 3D Gaussians to ground geometry and semantics, enabling manipulation policies to explicitly reason about spatial relationships.

Mohan Liu et al. · arXiv 2607.12356 source

TrustVLA: Defending VLA's "visual backdoors" at inference time · vla

End users can't audit a VLA's deployment pipeline; a poisoned model can behave normally on clean footage yet hijack long-horizon policies the moment a small visual trigger appears, before any failure surfaces. This work proposes a mechanism-guided inference-time defense that identifies and suppresses backdoor triggers without retraining—a safety angle rarely seen in VLA work but a critical one.

Pinhan Fu et al. · arXiv 2607.12571 source

Jetson-PI: Getting VLA to run real-time control on Jetson · vla

VLA is computationally heavy, causing high latency and low control frequency when deployed on low-power boards like the Jetson Orin. This work uses "lookahead-aligned asynchronous inference" to enable onboard real-time control, addressing a real bottleneck for on-device deployment on physical robots.

Zebin Yang et al. · arXiv 2607.12659 source

GaitSpan: A humanoid that can walk shouldn't have to learn to run from scratch · locomotion

Humanoid robots shouldn't have to relearn locomotion from zero just to jog or run. GaitSpan lets gait "grow" from walking to running, avoiding the skill fragmentation that comes from existing methods relying on preset gait timing or imitating motion clips.

Kwan-Yee Lin et al. · arXiv 2607.12114 source

Other papers today: ChunkFlow eliminates boundary jitter in chunked action heads (arXiv 2607.12992 source); Reducing Temporal Redundancy cuts VLA inference time redundancy for a speedup (arXiv 2607.12287 source); ExToken uses structured exploration to improve VLA reinforcement fine-tuning efficiency (arXiv 2607.12931 source); DenseReward learns dense rewards from synthesized failures (arXiv 2607.13033 source); Robust In-Hand Manipulation achieves robust in-hand manipulation without external perception (arXiv 2607.12105 source); Hy-Embodied-VLM-1.0, Tencent's efficient embodied VLM from its Hunyuan team (arXiv 2607.12894 source); MAMMOTH is a robust off-road end-to-end policy under missing modalities (arXiv 2607.12965 source); TerraZero is a zero-demonstration self-play driving simulator (arXiv 2607.13028 source).

Open Source · Tools · Benchmarks

· WorldArena 2.0 Challenge: a "real exam" for world models launched by Jiqizhixin (Chinese AI media outlet) and others; Track 2 places world models inside an online reinforcement-learning loop—receiving policy actions, predicting the next observation, and returning a reward—testing whether a model can "serve as a training environment" rather than just generate a video clip source (WeChat, CN)

· Ruka-v2: an open-source tendon-driven dexterous hand under $1,500, with an upgraded wrist and e-flesh soft fingertips, capable of writing, flipping pages, and holding a pen source (WeChat, CN)

II. Funding and Deals

Walden Robotics | First round out of stealth | $300 million | $1.1 billion valuation · industrial

Co-led by Toyota and Deviation Capital, with participation from Nvidia, Boeing, Samsung, Prologis, and CoreWeave, among others. The company spun out of the Toyota Research Institute (TRI) this past January; its founder and CEO is Russ Tedrake, an MIT professor and former head of TRI's large behavior models effort, pursuing a "full-stack physical AI" approach that builds on Diffusion Policy and Large Behavior Models. The most notable point isn't the valuation: its robots have been in production at a Toyota North America factory since February, going from pilot to production line in under two months, targeting labor-constrained industries including automotive, aerospace, semiconductors, electronics, logistics, and life sciences.Source: The AI Insider source

RoboParty | Angel++ round + Pre-A round | Nearly RMB 500 million · embodied

CATL led the Pre-A round exclusively, with Shunwei Capital, Matrix Partners China, and Xiaomi's strategic investment arm adding on; the two rounds together total nearly RMB 500 million. The company builds an open-source full-stack embodied platform, with a founding team of 22-year-olds from the Harbin Institute of Technology background; CATL leading the round in person signals that battery and industrial capital are pushing upstream into embodied platforms along the "hardware-components" chain.Source: Embodied Emergence (Chinese embodied-AI media) source (WeChat, CN)

Monumental (Netherlands) | New round | $32 million · industrial

Led by Khosla Ventures, founded by two former Palantir employees; the funds will go toward expanding its fleet of 150 brick-laying robots across Europe and into the United States, addressing labor shortages in construction trades in places like the UK.Source: Tech Funding News source

Cytronic | Seed round | $13.5 million · industrial

Reported exclusively by Axios; builds warehouse fulfillment (order picking) robots, targeting the mid-sized warehouse segment that large-scale automation hasn't yet reached.Source: Axios source

Yuejiang Technology | Upcoming ChiNext IPO review | First "H-to-A" listing in the Greater Bay Area · embodied

Yuejiang, a maker of collaborative and humanoid robots, goes before the review committee next week and will become the first company in the Guangdong-Hong Kong-Macao Greater Bay Area to complete an "H-to-A" listing (moving from a Hong Kong listing to an A-share listing).Source: 21jingji (Chinese financial media) source

Deyi Medical | Series A++ round | Tens of millions of RMB · embodied

A fund under Tiantu Capital invested exclusively; the company builds AI-plus-healthcare embodied robots and has now closed two funding rounds within the year.Source: Tiantu Capital source (WeChat, CN)

III. Commercialization and Deployment

Agibot's Genie G2 hits 15,000 units cumulative, launches G2 Max with JD Logistics · humanoid ⚠️ vendor claim

According to Agibot (Zhiyuan Robotics, Chinese humanoid startup), its Genie G2 reached its 15,000th unit off the line by late June and has been deployed on production lines at a Longqi Technology factory, taking roughly three months to go from 10,000 to 15,000 units; for reference, Omdia data put total global humanoid shipments last year at around 13,000 units, meaning this single batch already exceeds last year's global total—Agibot held nearly 40% of global humanoid market share last year. Its paired "Data Collection 2.0" dual-track collection (real robot plus simulation) aims to shore up sim-to-real gaps, and is also central to its partnership with Aishida—though it should be noted that Aishida assembles the Cooper quadruped, not the full humanoid unit. Today, JD Logistics and Agibot jointly launched the embodied robot Genie G2 Max. Leading in production volume is not the same as a "technical leap"—the mass-production numbers are, more than anything, favorable conditions for accumulating data.Source: Zhizhen Guan (Chinese robotics media) source (WeChat, CN)

Unitree's G1 goes abroad to "lay bricks": Japanese construction firm puts a humanoid on-site · humanoid

A Japanese construction company has deployed Unitree's G1 humanoid robot on a construction site—another instance of a Chinese humanoid robot going overseas into a real industrial setting, rather than a stage demonstration.Source: IT Home (Chinese tech media) source

Gausium named world's No. 1 commercial cleaning robot supplier by IDC · industrial

An IDC report ranks Gausium as the world's top commercial cleaning robot supplier by shipments; the company says it is shifting from winning flagship contracts toward scalable growth. Cleaning remains one of the few commercial service robot categories that has already achieved scale.Source: PR Newswire source

Boston Dynamics trials robot dog Spot for door-to-door package delivery · adjacent ⚠️ pilot

Boston Dynamics demonstrated its Spot robot completing "the last fifty feet" of door-to-door package delivery; this remains a pilot demonstration, not scaled delivery.Source: The Verge source

Warehouse and retail robots land at multiple sites: Amazon, Tesco, Colruyt · industrial

Amazon has reportedly been building a roughly $1 billion robotic warehouse in western Sydney; UK grocer Tesco is piloting a 6.5-foot-tall robot that patrols stores to automatically detect shelf out-of-stocks; Belgium's Colruyt has teamed up with KION to establish an AI warehouse robotics center. Warehousing and shelf fulfillment are becoming among the fastest-scaling deployment scenarios.Source: The Urban Developer source

IV. Industry Developments

LimX Dynamics releases COSA 0.5 humanoid brain system, reiterates "the model is not the brain" · humanoid

LimX Dynamics released a teleoperation-free, single-take demo of its humanoid brain system LimX COSA 0.5: the robot Oli continuously performs long-horizon household chores including hanging laundry, putting things away, stacking and moving boxes, and deep bending to pick up objects. Founder Zhang Wei states plainly that "the model is a skill, the system is the brain"—the brain should be an operating system that orchestrates VLM/LLM/VLA and other models and tools, not a single ever-growing model. The system has three layers: S2 cognitive scheduling (roughly 1 Hz), S1 skill layer (VLA/world action models generating whole-body motion policies), and S0 whole-body motor control (LimX WBT, roughly ten million parameters, running at 1000 Hz on the physical body). LimX reports WBT's average joint position error at 12.85 millimeters, slightly better than SONIC, the industry's strongest whole-body tracking model it benchmarks against (13.75 millimeters). This contrasts with Figure's emphasis on an end-to-end unified neural system, putting the debate over "is the brain a system or a large model" on the table; the single-take demo remains a capability showcase—scaling will depend on whether it can be reliably reproduced.Source: DeepTech (Chinese tech media) source (WeChat, CN)

Stardust Intelligence releases embodied foundation model Lumo-2 and physical AI agent Philia · world-model ⚠️ vendor claim

Lumo-2 is positioned as a "household Latent World-Action Model": before acting, it predicts task-relevant physical changes in a compressed latent space, then generates actions accordingly—sitting between traditional VLA and full video world models, and covering 22 household tasks such as catching a ball, flipping food in a pan, weighing rice, and tying a bow. According to its paper, end-to-end latency on a single RTX 5090 drops from about 253.66 ms to 93.53 ms, outperforming baselines such as π0.5 and Fast-WAM overall. Also released alongside it, Philia turns robots into "physical executors" that agents can call, exposing standardized capabilities via a Robot Gateway, pointing toward a "systems engineering" phase for embodied AI.Source: AI Faner (Chinese tech media) source (WeChat, CN)

XPeng's second-generation VLA passes localization certification in Germany, He Xiaopeng attends in person · autonomy ⚠️ certification claim

XPeng's Turing AI-assisted-driving NGP second-generation VLA model has completed localization certification testing in Europe, with He Xiaopeng traveling to Germany personally for the certification; the company states the same model adapts to European road signs and traffic regulations without retraining, aiming to be the first Chinese brand to deploy advanced driver-assistance in Europe. This is the company's own certification claim; commercial deployment still awaits local regulatory approval.Source: IT Home (Chinese tech media) source

Hyundai reportedly moving to buy out SoftBank's stake in Boston Dynamics · humanoid

According to Korean media reports, as a deadline approaches, Hyundai Motor is reportedly moving to acquire SoftBank's stake in Boston Dynamics, further consolidating ownership of the veteran robotics company.Source: Korea Financial News source

UBTech signs "the Mercedes-Benz of dealerships" group, humanoid robots enter auto dealership channels · humanoid

UBTech has signed a sales partnership for its humanoid robots with a dealership group known as "the Mercedes-Benz of 4S stores," which becomes its first authorized partner in the Beijing-Tianjin-Hebei region—robots are beginning to reach customers through auto dealership channels.Source: Sina Finance source

Policy and capital both ramp up ahead of WAIC, embodied AI gets its own track · adjacent

This year's World Artificial Intelligence Conference (WAIC 2026) is establishing a dedicated embodied AI zone for the first time, splitting "AI computing" and "embodied intelligence" into two independent core tracks; concurrently, Guangdong launched a RMB 100 billion strategic emerging industry fund, and robotics ETFs have seen consecutive days of inflows. On the official data side, China's National Bureau of Statistics reports 3D printing equipment output up 48.5% year-over-year in the first half of the year. Policy, capital, and industry are moving in sync, setting the tone for the sector right now.Source: Cailianshe (Chinese financial media) source

Former Qualcomm autonomous-driving VP set to join Nvidia, "Qualcomm alumni" and "XPeng alumni" converge · autonomy

Qualcomm's former VP of engineering for autonomous driving is set to join Nvidia's autonomous driving team, seen as another reinforcement of Nvidia's automotive autonomous driving effort.Source: Sina Finance source

Hardware · Supply Chain

· Galaxea (Chinese robotics startup): will showcase at WAIC a dexterous-hand solution priced at the RMB 10,000-level (⚠️ vendor claim, conference promotion) source

· Linker Hand: claims peak monthly output of over 4,000 units and over 80% global share of high-degrees-of-freedom dexterous hands, with a valuation of RMB 40 billion (⚠️ vendor claim) source (WeChat, CN)

· Cost breakdown (Yiou, Chinese tech media): a 17-actuator, 5-sensor linkage-driven dexterous hand has a BOM cost of about RMB 51,800, with actuation and transmission accounting for roughly 60% of it—this is the reference point against which the "RMB 10,000-level" pricing claim should be read source (WeChat, CN)

· Horizon Robotics' Sunrise S600: has entered mass-production validation for its embodied AI chip, claiming recognition from 20+ leading customers (⚠️ mass-production validation claim) source

· Joyson Electronics: will showcase a full robotics component lineup at WAIC—dexterous hands, head assemblies, and embodied AI "brains" source

Top comments (0)