Today's Highlights

· Bezos's Physical AI company Prometheus closes ~$12B Series B at ~$41B valuation, targeting a "universal engineer" AI for designing and manufacturing physical products.

· Qianxun Intelligence locks in 1.5B RMB Series A+, raising nearly 5B in four rounds over three months; Yunfeng (Jack Ma) and Shunwei (Lei Jun) join, with self-developed VLA model Spirit series already running on CATL production lines.

· World models move from conference halls to engineering: 1X establishes World Model Lab; Hugging Face LeRobot integrates the first world model VLA-JEPA (fine-tunable with just 13 trajectories).

· MotionWAM delivers real-time "world-action model"-driven humanoid whole-body loco-manipulation using a monocular camera, reportedly outperforming VLA baselines by 30% on Unitree G1.

· Policy and commercialization accelerate: MIIT and SASAC launch the 2026 Humanoid Robot Real-Scene Training Initiative; Shenzhen opens city-wide Robotaxi testing and commercial operation from July 1.

MotionWAM: A Foundation World-Action Model for Real-Time Humanoid Loco-Manipulation · locomotion

This work brings the "world model" out of passive video prediction and into a real-time whole-body control loop for humanoids — driving full-body mobile manipulation with a monocular egocentric camera. It is one of the rare end-to-end real-robot results on the WAM (world-action model) path.

arXiv 2606.09215 https://arxiv.org/abs/2606.09215 · Analysis: 具身智能情报官 source

The method conditions the policy directly on intermediate denoising features of a video world model rather than generating complete future frames. Training proceeds in three stages: first, future-frame generation pretraining on ~2,136 hours of human egocentric and humanoid robot video; next, cross-embodiment action post-training on heterogeneous Unitree G1 data (varying end-effectors and action annotation formats) using a single joint flow-matching objective that updates both the video and action networks simultaneously; finally, target-task fine-tuning on whole-body teleoperation demonstrations. The analysis reports that the system outperforms VLA baselines on whole-body mobile manipulation on the Unitree G1 by more than 30%, and can walk and manipulate simultaneously rather than learning locomotion and manipulation separately.

VLA-JEPA: Injecting JEPA World Modeling into Vision-Language-Action Models, Fine-Tunable with 13 Trajectories · world-model

Hugging Face adopts this as the first world model integrated into LeRobot. Its significance lies not in a single SOTA result but in packaging "world model + extremely low-sample fine-tuning + consumer GPU inference" into an open-source, reproducible engineering baseline that directly lowers the barrier to real-robot learning.

Hugging Face LeRobot · Analysis: 真机智能 source

The model learns not only "what action to take" from observations, but also the action-relevant environmental dynamics via a JEPA world model. During training, a world modeling objective is added to the predictor by conditioning on V-JEPA2, enabling pretraining on human video; during inference, the world model is discarded entirely, leaving only the standard VLA structure (Qwen backbone + action head), balancing capability and deployment cost. The authors report that fine-tuning on a new task requires only ~13 trajectories and that inference runs on low-cost GPUs. This "use world model for training, keep only policy for inference" approach aligns with LeRobot's recent direction of treating world models as first-class citizens.

MemoryVLA: Curing VLA "Goldfish Memory" with Perceptual-Cognitive Memory · vla

This work targets a structural weakness of standard VLAs — each inference step sees only the current frame with no memory of what happened moments before — causing frequent failures on long-horizon multi-step tasks like "put all cups on the table into the cabinet."

ICLR 2026 · Analysis: AI共享前沿 source

The work introduces two layers of memory — perceptual and cognitive — into VLAs, allowing the model to retain and recall historical information across long-horizon tasks (e.g., completed/incomplete sub-goals, past failures) rather than "forgetting" frame by frame. The analysis focuses on the practical deployability from principles to local deployment.

Prometheus (Bezos) | Series B | ~$12B | Valuation ~$41B · adjacent

Co-led by Amazon founder Jeff Bezos and Stanford professor Vik Bajaj, operating independently of Amazon and Blue Origin with a team of ~150. Investors in this round include JPMorgan, Goldman Sachs, BlackRock, DST Global, Arch Venture, and Bezos himself. The company debuted in November 2025 with $6.2B and has now raised over $18B in total over seven months. This is not a robotics company — it aims to be a "universal engineer" AI for the physical world, automating the design and manufacturing of complex physical systems (from jet engines to drug molecules) targeting aerospace, automotive, and semiconductors. This is among the largest single bets in the current Physical AI capital wave, and the direction extends from "building embodiments" to "using AI to rewrite engineering design itself."Source: Tech Times https://www.techtimes.com/articles/318330/20260613/prometheus-ai-startup-bezos-raises-12-billion-41-billion-valuation.htm (also per TechCrunch, Axios)

Qianxun Intelligence Spirit AI | Series A+ | 1.5B RMB · humanoid

This round was joined by Yunfeng Fund (Jack Ma) and Shunwei Capital (Lei Jun), marking the company's completion of four rounds in three months and cumulative fundraising of nearly 5B RMB — setting a new funding velocity record in the embodied AI sector. Founded in January 2024 in Hangzhou by former Rokae Robotics co-founder and CTO Han Fengtao and Tsinghua Institute for Interdisciplinary Information Sciences assistant professor Gao Yang, the company pursues a full-stack self-developed hardware and software approach: on the software side, a self-developed VLA foundation model series called Spirit; on the hardware side, a humanoid called Moz1 with 26 degrees of freedom and integrated force-controlled joints. Products are already running stably on CATL battery production lines, with cooperation with JD.com in retail scenarios. The company claims its Spirit v1.6 tops the NVIDIA-co-initiated RoboArena leaderboard, surpassing Cosmos3 and Physical Intelligence's Pi0.5 — this result is the company's own claim and should be treated as a capability reference rather than an independent determination.Source: 创新创业科学家 source; 36Kr https://m.36kr.com/p/3837372018362885

Zhuxian Technology | Files for Hong Kong IPO · autonomy

This L4 autonomous trucking solution provider has initiated the Hong Kong listing process, with shareholders including NIO Capital and iFlytek. Its prospectus shows annual revenue of ~340M RMB and a net loss of ~170M RMB — consistent with the rest of the sector, where scaled revenue and profitability remain unsolved problems. Autonomous long-haul logistics is rushing to capital markets this year, another stress test of the sector's shift from "test mileage" to "commercial delivery."Source: 集微网 https://www.laoyaoba.com/n/1050130; 新浪财经 https://m.sohu.com/a/1035899242_430392?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334

Rekise Marine | $9.7M · autonomy

This Indian autonomous maritime platform company raised $9.7M to accelerate development of autonomous naval and marine platforms. Maritime autonomy is a relatively niche but strongly dual-use (civilian and military) segment of embodied autonomy, with capital beginning to extend toward "surface and subsurface autonomy."Source: 搜狐 https://m.sohu.com/a/1036153915_506163?scm=10001.325_13-325_13.0.0-0-0-0-0.5_1334

Juzhi Home | Angel Round · adjacent

Positioning itself as "China's first embodied robot aggregated information platform," this round was invested by the family office of manufacturing giant Lens Technology. Funds will be used to deepen online and offline presence and build out a full-industry-chain service ecosystem. Primary market enthusiasm spilling over into "shovel sellers / platform plays" at the service layer is a companion phenomenon to the sector entering a small-scale deployment phase.Source: 投资家 source; 具身之家 source

Shenzhen Opens City-Wide Robotaxi Testing and Commercial Operation from July 1 · autonomy

The revised Shenzhen Special Economic Zone Intelligent Connected Vehicle Management Regulations take effect July 1, 2026, orderly opening the entire city to road testing and commercial operation of driverless taxis. The shift from designated test zones to "city-wide opening + commercialization" is a key policy unlock pushing Robotaxi from pilot to normalized operation in a tier-1 city, and provides institutional space for domestic players to expand their operational mileage.Source: 百姓大事小情 source

Tesla Applies to Deploy Up to 5,000 Robotaxis in Las Vegas · autonomy ⚠️ Application pending approval

According to documents filed with Nevada's Transportation Authority (NTA), Tesla plans to deploy up to 5,000 driverless taxis in Clark County — still in the application stage and not yet approved. Meanwhile, official Texas records show Tesla's autonomous ride-hailing fleet remains far behind Waymo in scale, and overseas media assessments of its Robotaxi progress are divided. The gap between ambitious planning figures and actual operating scale is something to distinguish carefully when assessing Tesla's Robotaxi deployment pace.Source: 维加斯资讯网 source; MSN https://www.msn.com/en-us/money/companies/official-texas-records-show-tesla-s-autonomous-ride-fleet-is-trailing-far-behind-waymo/ss-AA25iGLV

PepsiCo and Gatik Expand Autonomous Freight Partnership to Three States · autonomy

PepsiCo and autonomous mid-range freight company Gatik are expanding their partnership to Texas, Arizona, and Arkansas, covering more fixed-route B2B short-haul transport. Compared to Robotaxi, the "fixed routes + large anchor customers" model for autonomous freight has a clearer commercial loop and is one of the higher-validation scenarios in autonomous driving today.Source: Food Business ME&A https://www.foodbusinessmea.com/pepsico-gatik-expand-autonomous-freight-partnership-across-texas-arizona-arkansas/

Hyundai Deploys 4 Spot Robot Dogs at the 2026 FIFA World Cup for Inspection and Security · industrial

Hyundai (via Boston Dynamics) is deploying 4 Spot quadruped robots at the 2026 FIFA World Cup for venue patrol, inspection, and crowd monitoring — the first appearance of this type of robot at a World Cup. This is a typical "established platform + major event security" deployment; inspection and security remain one of the few commercial use cases where quadruped robots have achieved repeated scale deployments.Source: CPG https://en.clickpetroleoegas.com.br/hyundai-took-four-spot-robot-dogs-to-the-2026-world-cup-which-will-patrol-inspect-and-monitor-crowds-at-tournament-venues-in-the-united-stat-mhbb01/

UBTECH's Ultra-Bionic Companion Robot U1: Nearly 4,000 Pre-Orders in 10 Days (Previously Reported; Figure Updated) · humanoid

Deposit is 3,000 RMB/unit (refundable), expected to launch June 30, available to adults only; the pre-order figure has risen from ~2,700 units in 6 days to nearly 4,000 in 10 days. Note: these are pre-order deposits, not delivery volumes — final installed units depend on final payment and shipping, and the industry is also debating the ethical boundaries of "nurturing-style" emotional companionship.Source: 大河报 source

1X Establishes World Model Lab, Betting on Large-Scale Embodied World Model Pretraining · world-model

Home humanoid company 1X has set up a World Model Lab focused on frontier research, with the goal of pursuing fully autonomous humanoid systems through large-scale embodied world model pretraining and building the foundation and reasoning models underpinning spatial intelligence. The company explicitly identifies "data" as the only enduring moat in embodied AI — the ability to collect, annotate, scale, and train on the right mix of web-scale data, human data, simulation data, teleoperation data, and robot self-generated data. Early-shipped NEO units will serve as remote data collection nodes, feeding real-world long-tail edge cases back and continuously updating the foundation model via OTA. Organizationally, former Luma AI founding researcher Sam Sinha has joined as Head of World Models. This echoes the same-day LeRobot integration of VLA-JEPA — world models are rapidly moving from papers and conference halls into the engineering mainstream at leading embodied AI companies.Source: 1X https://www.1x.tech/discover/1x-world-model-lab

Xingyuanzhi Releases Embodied Interactive World Model ω-EVA, Claims It Closes the Loop Between World Models and Action Decisions · world-model

ω-EVA's pitch is that the world model can evaluate the "latent consequences" of candidate actions in real time within the same decision cycle and feed them back to the action generation module, completing a "candidate action → imagined consequence → correction" loop that bridges embodied world models to real control. On the company side, there is also news of the "embodied brain" raising ~1B RMB over ten months. Whether this loop can run stably on real hardware remains to be seen, but the direction is consistent with how 1X and LeRobot are incorporating world models into control.Source: 星源智 source; 21财经 https://m.21jingji.com/article/20260613/66285795a9b477857bb41fd759eb1689.html

BAAI Conference Day Two: Embodied Intelligence Features as a Standalone Core Topic for the First Time · embodied

Following the previous day when world models dominated the agenda (Wujie·Physis-v0.1, Fei-Fei Li clarifying the definition of "world models" — both previously reported), the June 13 Embodied Intelligence and Humanoid Robot Forum saw multiple team leads including Galaxy General Wang He take the stage, with live demonstrations of autonomous operations such as flower arranging and ping-pong; BAAI also outlined four categories of world model technical approaches. The signal: embodied intelligence has been elevated from "a breakout session at an AI conference" to a standalone core topic.Source: 人工智能纵横 source; 智东西 source

Daxiao Robotics' "Kaiwu" World Model Claims Top Spot at CVPR 2026 WorldArena · world-model ⚠️ Vendor claim

SenseTime-affiliated Daxiao Robotics claims its "Kaiwu" world model achieved first place across multiple leaderboards at CVPR 2026 WorldArena with a small parameter count (the official account gives inconsistent figures between 2B–4B), surpassing Sora and Veo. The WorldArena evaluation benchmark itself is real (Tsinghua team, arXiv 2602.08971 https://arxiv.org/abs/2602.08971), but public channels have not been able to independently verify the specific rankings behind "Kaiwu topping / four leaderboard firsts" — the claims are currently vendor-only and should be treated as capability assertions, not verified findings.Source: AI智件 source

JD.com Partners with RoboNow to Build End-to-End Robot Service Ecosystem · adjacent

JD.com and RoboNow are building a robot service chain covering deployment, operations, and after-service. A platform e-commerce giant integrating "robot-as-a-service" into its existing fulfillment and channel infrastructure is another example of a major player shifting from buyer/venue operator to ecosystem builder.Source: Gasgoo https://autonews.gasgoo.com/articles/icv/jdcom-partners-with-robonow-to-build-end-to-end-robot-service-ecosystem-2065716997013479424

Fenglong Shares Board Renewal: UBTECH Founder Zhou Jian Elected Chairman · humanoid

Listed company Fenglong Shares has completed its board renewal, with UBTECH founder Zhou Jian elected as chairman. A humanoid complete-machine manufacturer's founder entering the governance layer of an A-share components and manufacturing company is a signal of deepening integration between industrial capital and listed companies around "robotization."Source: 每日经济新闻 https://www.nbd.com.cn/articles/2026-06-12/4425314.html