Today's Highlights
· Shenzhen-based AI²Robotics (Chinese humanoid startup) and X Square Robot (Chinese embodied AI startup) both crossed the 20 billion RMB valuation mark on the same day: AI²Robotics raised nearly 5 billion RMB in a new round, while X Square Robot completed four rounds in two months with a reported valuation exceeding 20 billion RMB — bringing China's embodied AI "billion-yuan club" to approximately 25 companies.
· Volkswagen is reportedly moving to terminate its autonomous driving partnership with Bosch — after roughly €1.5 billion invested over four years, an internal assessment deemed the technology "not competitive," and the company is pivoting to external suppliers with a target of signing a new contract before September.
· South Korea has elevated physical AI to a national strategy: the government plans to invest 20 trillion KRW in manufacturing AI, commercialize industry-specific humanoid robots across 10 sectors by 2028, and has joined 11 companies to establish a Physical AI Industry Association.
· Tesla is rolling out FSD 14 Lite to owners of older HW3 vehicles, though those vehicles remain ineligible to join Tesla's robotaxi fleet.
· Multiple world model announcements landed on the same day: Xpeng unveiled predictive world model framework X-Mind at CVPR 2026 (in service of VLA 2.0), while Wujie Power (Chinese robotics startup) released its "latent-space world model MWA," claiming top benchmark rankings (⚠️ vendor claim).
I. Research Papers
PhysisForcing: A Physics-Constrained Reinforcement Approach for Robotic Manipulation World Simulators · world-model
Addresses a core limitation of video-generative world models for embodied manipulation: both general-purpose video generators and robot-data fine-tuned models produce physically implausible outputs — discontinuous trajectories, interpenetration between robots and objects — undermining their reliability as world simulators. This work injects physical consistency into the generation process via reinforcement, improving the physical plausibility of manipulation rollouts. HF↑30 on the day of release.
Peiwen Zhang et al. · arXiv 2606.28128 source
Using "Translation" as a Bridging Action: Transferring Human Manipulation Skills to Bimanual Robots · manipulation
Human motion data is cheap, abundant, and diverse — the most promising resource for scaling robot learning — but the morphological gap between human hands and parallel grippers makes direct transfer difficult. This work uses a class of "translation" bridging actions to map human demonstrations onto bimanual robots with parallel grippers, demonstrating that new manipulation skills can be learned from human motion data. HF↑27 on the day of release.
Sijin Chen et al. · arXiv 2606.28133 source
Injecting a Single Grounded 3D Anchor Point into the Action Head Unlocks Spatial and Task Generalization in VLAs · vla
VLA models achieve flexible manipulation through large-scale vision-language pretraining, but remain brittle along two axes at test time: object positions outside the training distribution (spatial generalization) and familiar scenes paired with new instructions (task generalization). This work finds that simply feeding a single 3D-grounded spatial point directly into the action head simultaneously improves both types of generalization — a minimal intervention with targeted effect.
Shiang-Feng Tsai et al. · arXiv 2606.27663 source
S²-VLA: State-Space-Guided VLA for Long-Horizon Manipulation · vla
Existing VLAs commonly suffer from perceptual drift and intent loss in long-horizon tasks due to accumulating errors. S²-VLA uses state-space-guided adaptive attention to suppress error accumulation; the team reports that with 2B parameters it outperforms 7B-scale baselines on multiple long-horizon manipulation benchmarks.
Zhipeng Xie et al. (Shanghai Jiao Tong University, East China Normal University) · arXiv 2606.27872 source · Commentary: Embodied Intelligence Notes source (WeChat, CN)
Drop and Recover: How Redundant Are VLA Models? · benchmark
VLAs directly inherit the oversized language backbones of pretrained VLMs — capacity far exceeding what is needed to process short robot instructions. This study uses a "drop-and-recover" probe to systematically measure how much model capacity closed-loop control actually requires, providing quantitative grounding for VLA pruning and compression with direct relevance to on-device deployment.
Guoheng Sun et al. · arXiv 2606.27755 source
SceneBot: Contact-Prompted Whole-Body Humanoid Tracking and Scene Interaction · locomotion
Current humanoid RL policies handle free-space motion well but struggle with contact-rich tasks — pure kinematic tracking cannot resolve the physical ambiguities that arise when interacting with objects or uneven terrain. SceneBot uses contact prompts to unify free-space walking, terrain traversal, and whole-body manipulation within a single motion-tracking framework.
Sirui Chen et al. (Guanya Shi group) · arXiv 2606.27581 source
Perceptual 3D Simulation: Predicting Scene Evolution via Physical World Modeling · world-model
"Given an image, predict how a scene evolves under a 3D transformation" is a shared goal across vision, graphics, and robotics — yet real systems, unlike ideal simulators, cannot fully access 3D geometry and dynamics, and must rely on partial, incomplete perceptual inputs. This work models physical evolution under perceptual constraints, advancing toward actionable, reality-level simulation.
Wanhee Lee et al. (Stanford, Daniel Yamins group) · arXiv 2606.27575 source
Other papers today: SpikeVLA (low-power real-time VLA using spiking neural networks, arXiv 2606.27807 source); Support-Constrained RL (improving real-world policies without real-robot experience, arXiv 2606.27475 source); Unleashing Infinite Motion (amplifying expressive quadruped locomotion with generative video priors, arXiv 2606.28237 source); DexCompose (reusing dexterous manipulation policies for single-hand multi-task execution, arXiv 2606.28323 source); Booster Lab (data-centric pipeline for deployable humanoid locomotion, arXiv 2606.27813 source); DIM-WAM (world-action modeling with diverse historical event memory, arXiv 2606.27677 source); ReWorld (learning better representations for world-action models in autonomous driving, arXiv 2606.27504 source).
Open Source · Tools · Benchmarks
· MemoBench: A world-modeling benchmark focused on dynamically changing environments — specifically testing memory consistency when targets move out of view or are occluded while the scene continues to evolve, filling a gap left by most benchmarks that only test targets within the field of view (arXiv 2606.27537 source).
· SimFoundry: A modular, automated scene generation system capable of zero-shot real-to-sim reconstruction from a single video, for use in policy learning and evaluation (arXiv 2606.28276 source, HF↑6).
II. Funding & Deals
X Square Robot (Chinese embodied AI startup) ｜ B+/B++/C Rounds ｜ Post-money valuation exceeding 20 billion RMB (~$2.8B) · embodied
Three consecutive rounds — B+, B++, and C — all closed within just over two months, making it the first company in the Greater Bay Area to reach a 20 billion RMB valuation. The investor lineup is an unusually complete roster of Chinese internet majors: Meituan, Alibaba, ByteDance, and Xiaomi each led different rounds, with Sequoia China, IDG, Source Code Capital, the National AI Industry Fund, and China Mobile also participating. The company focuses on embodied foundation models and integrated robots, with deployments already live in real-world settings including 58 Daojia (Chinese home services platform) and a component production line at a German luxury automaker. Source: Pencil News source (WeChat, CN)
AI²Robotics (Chinese humanoid startup) ｜ New Round ｜ ~5 billion RMB (~$700M) ｜ Valuation exceeding 20 billion RMB · embodied
Announced its 20 billion RMB valuation on the same day as X Square Robot. This round was jointly led by the National SME System Fund, the Guangdong Provincial AI Fund, Shenzhen Capital Group (深创投), Nanshan Strategic Emerging Industry Investment, and strategic investors including Moutai Group (Chinese spirits conglomerate), China Merchants Capital, and Sino Biopharmaceutical — with the valuation doubling from the ~10 billion RMB Series B in February. The company positions itself around general-purpose productivity robots led by AlphaBot2 and a proprietary brain-inspired VLA, with its own production line now at hundreds of units per month and a target of reaching tens of thousands this year. Source: RecodeX source (WeChat, CN)
SpeedBot (Chinese industrial robot vision startup) ｜ Series B++ ｜ 100M+ RMB range · industrial
This round was co-invested by the National SME Hunan Sub-Fund managed by Dachen Caizhì (Chinese PE firm) and Changsha Urban Development. The company has previously completed seven rounds, including a Series B of 300 million RMB. SpeedBot uses an industrial world model to build physical cognition for manufacturing scenarios, supporting embodied AI deployment on production lines. Source: Changsha Enterprise Listing Service Center source (WeChat, CN)
Delta Intelligence ｜ Angel+ Round ｜ Undisclosed · world-model
A humanoid robot foundation model company, this round was backed by Zhuhai Science and Technology Industry Group, Haiwan Capital (Chinese VC), Meihua Ventures (Chinese VC), semiconductor investor Jiaoyuan Assets, and a strategic automotive industry participant; existing investor Huaying Capital (Chinese VC) followed on. Source: Tech Capital Circle source (WeChat, CN)
Proception ｜ New Round ｜ $11 million · embodied
A Y Combinator-backed robot dexterous hand startup whose founder Jay Li previously served as technical lead for Tesla's Optimus program. The company announced this funding round simultaneously with the settlement of a trade secret lawsuit brought by Tesla (Tesla withdrew the suit this month). Source: TechCrunch source
Archon Robotics ｜ Seed Round ｜ Hundreds of millions of RMB · embodied
Positioned as a general-purpose whole-body embodied intelligence developer, the company has completed a seed round in the hundreds of millions of RMB. Source: Gasgoo source
Qingyu Technology ｜ Pre-Series A ｜ Undisclosed · embodied
A portfolio company of Shunwei Capital (Chinese VC fund backed by Xiaomi's Lei Jun), focused on tendon-driven soft robots — a direction widely regarded as one of the most difficult embodied robotics approaches to bring to production. The company claims its first-generation bionic flexible product is among the few tendon-driven systems globally to reach mass production. Source: Shunwei Capital source (WeChat, CN)
III. Commercial Deployment
Shenzhen Further Deregulates Robotaxi Commercialization · autonomy
Reports indicate Shenzhen will advance the legalization of commercial robotaxi operations starting July 1. As of 2026, 26 cities across China have been approved to operate driverless robotaxi services with commercial fares, with several jurisdictions progressively removing the requirement for a safety operator in the vehicle. Industry discussion has shifted from "can the technology operate" to questions of operations and per-vehicle unit economics. Source: AI Dimension Bureau source (WeChat, CN)
Aurora's Autonomous Trucking Safety Case Endorsed by Third-Party Audit · autonomy
An independent third-party audit has endorsed Aurora's autonomous trucking safety case — a materially more credible safety signal than vendor self-assessment, and a substantive validation for the company's plans to scale commercial driverless long-haul trucking operations. Source: Truck News source
Faraday Future (FF) Reports 242 Robot Units Shipped, March–June · humanoid ⚠️ target figure
FF reports that its robotics division shipped 242 units between March and June, ahead of expectations; Jia Yueting (FF's founder) has raised the full-year shipment target to 2,000 units accordingly. The 242 units represent actual deliveries; the 2,000-unit figure is a forward-looking target, not achieved capacity. Source: Sohu source
UBTECH Reports 5,000+ Pre-orders in 20 Days for New Humanoid Robot · humanoid ⚠️ pre-order figure
UBTECH (Chinese humanoid robot company) reports that its new life-like humanoid robot surpassed 5,000 pre-orders within 20 days, with the product formally unveiled the following day. Pre-orders are not deliveries; actual production capacity and fulfillment remain to be verified after the official launch. Source: MyDrivers source
Pudu Robotics Plans "Fully Robot-Staffed Hotel" with Shenzhen CITD · industrial ⚠️ planned
Pudu Robotics (Chinese service robot company) announced plans to build what it calls the world's first fully robot-operated hotel in Guangdong, China, targeting an opening in 2027. This is a planning-stage announcement, not a completed or operational deployment. Source: Let's Data Science source
IV. Industry Developments
Volkswagen Plans to Terminate Autonomous Driving Partnership with Bosch · autonomy
Volkswagen plans to exit its 2022 joint development agreement with Bosch on driver assistance and autonomous driving systems. After a cumulative investment of roughly €1.5 billion over four years, an internal assessment concluded the technology was "not yet competitive." Volkswagen intends to terminate the contract under its terms and pivot to directly procuring software and hardware from external suppliers, targeting a new engineering supply contract before September — a move that runs in parallel with a sweeping cost-reduction program that includes up to 100,000 job cuts and the closure of four factories in Germany. Notably, Volkswagen has taken a different approach in the Chinese market, pursuing local intelligent driving partnerships there, with the two regional strategies now visibly diverging. Source: Autohome source
South Korea Elevates Physical AI to National Strategy · adjacent
The South Korean government has released a comprehensive physical AI package: planned investment of 20 trillion KRW in manufacturing AI alongside a "National AI Data Repository" to preserve skilled craftspeople's expertise in digital form; a target to commercialize industry-specific humanoid robots across 10 sectors by 2028; and a goal to grow the robotics market from 5.6 trillion KRW in 2021 to over 20 trillion KRW by 2030, with one million robots deployed, and an ambition to become the global leader in physical AI by 2030. On the industry side, Korea Startup Forum and 11 companies jointly established a "Physical AI Industry Association," while Samsung and others announced major investments at sites including Gumi. The 2028/2030 targets are forward-looking government commitments, though the policy announcements and industry organization activity represent substantive real actions rather than single-vendor claims. Source: Seoul Economic Daily source
Xpeng Unveils Predictive World Model Framework X-Mind · world-model ⚠️ vendor claim
Xpeng (Chinese EV and autonomous driving company) presented its world model roadmap at the CVPR 2026 Embodied Foundation Model Deployment Workshop, releasing X-Mind — a predictive world model framework enabling vehicles to "simulate the future before deciding," building on the prior X-World, X-Foresight, and X-Cache lineage. The framework emphasizes proactive reasoning, controllable generation, and long-horizon prediction to support the next-generation VLA 2.0; a technical report is forthcoming. This is a capability roadmap announcement; production-level results remain to be validated. Source: XPENG source
Wujie Power Releases Latent-Space World Model MWA · world-model ⚠️ vendor claim
Following a $200M+ angel round last month, Wujie Power (Chinese embodied AI startup) has released MWA™, which it describes as the world's first "long-horizon bidirectional physical causal chain" latent-space world model. The company claims it abandons the next-frame video prediction paradigm in favor of modeling physical causality directly in latent space, and asserts it tops a leading embodied intelligence benchmark. "World's first" and "top-ranked" are vendor claims that require independent verification. Source: Jiqizhixin (Chinese AI media) source (WeChat, CN)
Zhiyuan Robotics (Chinese humanoid startup) Takes Stake in Force Sensor Developer Lanpoint Touch · hardware
Following its previously reported milestone of the 15,000th unit rolling off its production line, Zhiyuan Robotics has taken a stake in force sensor developer Lanpoint Touch (Chinese sensor startup), extending its supply chain upstream into six-axis force and tactile sensing to strengthen its standardized supply chain ecosystem. Source: Guandian.cn source
Deep Robotics Rebrands and Announces "1+X+N" Global Strategy · embodied
Hangzhou-based quadruped robot maker "Hangzhou Deep Robotics" has dropped the city name from its branding, rebranding simply as "Deep Robotics," and unveiled a "1+X+N" strategy at its Global Partner Summit to accelerate international expansion of commercial embodied intelligence. Source: Yahoo Finance source
Meta Reported to Acquire Robotics AI Company to Accelerate Humanoid Development · adjacent ⚠️ unconfirmed report
Reports indicate Meta has acquired a robotics AI company to accelerate its humanoid robot development; transaction details and the identity of the acquisition target have not been officially confirmed. Source: Stocktwits source
Hardware · Supply Chain
· Ruiyan Intelligent Control (Chinese dexterous hand supplier): Reported to be the exclusive dexterous hand supplier for dozens of Chinese embodied large-model companies, emerging as a hidden chokepoint as the dexterous hand segment scales (China's dexterous hand shipments are projected to grow from 19,200 units in 2025 to approximately 70,000 units in 2026). source (WeChat, CN)
· 3D Force-Tactile Sensors: A team from South China University of Technology has developed multiple dexterous hand designs incorporating a proprietary low-cost, high-sensitivity 3D force-tactile sensor, targeting cost reduction in manufacturing — tactile sensing is a key bottleneck in moving dexterous hands from "mechanical grasping" to "intelligent manipulation." source (WeChat, CN)
Top comments (0)