Today's Highlights
· Uber and Waymo end nearly three years of Robotaxi partnership in Phoenix: Waymo's fleet shifts to DoorDash delivery and Via transit, while Uber pivots to Lucid and Nuro with plans to deploy 20,000 vehicles over six years; both companies continue their partnership in Austin and Atlanta.
· UBTECH Robotics (Chinese humanoid robot company) launches full-size ultra-bionic humanoid robot U1 under consumer brand "UWORLD," priced at RMB 119,800–990,000 with 88 degrees of freedom, targeting home emotional companionship; pre-sale orders exceed 11,000 units (⚠️ pre-sale figures).
· World model startup funding wave continues: Forbes reports the sector has raised billions of dollars this year; today Tsinghua-affiliated Liqi Intelligence closed a seed round of hundreds of millions of RMB, while Dexmal completed a RMB 1 billion Series B targeting an IPO (valuation inherited from Zibianliang and Zhipingfang at RMB 20 billion).
· Apptronik opens an approximately 90,000 sq ft "Robot Park" training facility in Austin and unveils Apollo 2 (bipedal + wheeled), partnering with Google DeepMind to collect real-world data for training humanoid AI.
· Laifual Drive (Chinese harmonic drive manufacturer) lists on the Hong Kong Stock Exchange as the first "harmonic reducer stock" on the Hong Kong market, raising approximately HKD 1.07 billion; one of only two Chinese manufacturers to have achieved mass production of harmonic reducers for humanoid robots (⚠️ three consecutive years of losses).
I. Research Papers
DreamForge-World 0.1: A Low-Compute, Real-Time Interactive World Model Preview · world-model
World models are typically constrained by offline generation and high compute demands. This paper delivers a controllable real-time interactive world simulation in a low-compute preview: it adds a residual action pathway on top of the LongLive autoregressive video stack (derived from Wan2.1-T2V-1.3B), drawing on Matrix-Game concepts to achieve "feed action, generate future frames simultaneously." It ranks as today's top community paper (HF 7↑), representing the engineering direction of world models shifting from offline long-video generation toward interactive, low-latency systems.
Daniyel Ayupov et al. · arXiv 2606.30292 source
Orca: Unifying Multimodal World Signals into a "Universal World Foundation Model" · world-model
Current world models tend to be built separately per task or modality. Orca attempts to learn a unified world latent space, acquiring shared representations from multimodal world signals and exposing them via multimodal readout interfaces — pointing toward a world foundation model paradigm of "one base, many downstream uses."
Yihao Wang et al. · arXiv 2606.30534 source
Heterogeneous Tactile Transformer: One Representation to Bridge Heterogeneous Tactile Sensors · perception
The biggest pain point in tactile data is that sensors are incompatible — a model trained on one sensor fails when transferred to another, making large-scale data aggregation difficult. HTT learns shared representations across heterogeneous tactile sensors, allowing contact data from different tactile hardware to be pooled for training. This is the algorithmic counterpart to today's wave of tactile data infrastructure investment (Qianjue, South China University of Technology, Weitai).
Jianxin Bi et al. · arXiv 2606.29948 source
Sim-to-Real Physical Modeling for Professional-Level Robot Table Tennis · manipulation
High-speed spinning table tennis trajectories are counterintuitive; a robot must track and return the ball precisely within fractions of a second, while real-environment training is both expensive and hazardous. The Sony AI team builds a high-fidelity physical model, trains reinforcement learning policies in simulation, then transfers to physical robots, targeting professional-level high-speed gameplay — a hard sim-to-real benchmark for dynamic, contact-constrained control tasks.
Christian Conti et al. (Sony AI) · arXiv 2606.28805 source
Human2Any: Constraint-Aware Compositional Transfer from Human to Robot · manipulation
Human video is a scalable source of manipulation supervision, but morphology gaps, scene variation, and robot feasibility constraints make direct transfer difficult. Human2Any uses constraint-aware compositional planning to translate human demonstrations into actions executable by different robots, mitigating the embodiment gap; co-authors include NVIDIA researchers.
Shuo Cheng et al. (incl. NVIDIA) · arXiv 2606.28813 source
AnyBody: Whole-Body Humanoid Control Driven by Any Keypoint Subset · locomotion
AnyBody proposes a unified whole-body humanoid controller that can be commanded at deployment time by any subset of body keypoints — specifying only hand, foot, or torso targets is sufficient to generate coordinated whole-body motion, improving flexibility for teleoperation and motion retargeting.
Shuning Li et al. · arXiv 2606.29209 source
Trust Your Instincts: VLA with Test-Time RL Using Its Own Confidence · vla
Applying reinforcement learning to vision-language-action models typically requires external environment feedback and predefined success signals. This paper lets the model use its own confidence as an intrinsic signal for test-time RL, enabling online self-improvement without external success judgments — lowering the deployment barrier for VLA reinforcement learning.
Siyao Chen et al. · arXiv 2606.29892 source
You Only Touch Once: 6-DoF Object Pose Estimation from a Single Contact · perception
Visual pose estimation often fails with occlusion, reflective surfaces, or transparent objects. YOTO recovers the full 6-DoF pose of an object from just a single simultaneous contact pair, without requiring contact history, providing a purely tactile pose source for contact-rich manipulation.
Pengfei Ye et al. · arXiv 2606.28899 source
Other papers today: Flow Matching in Feature Space for Stochastic World Modeling (authors include INRIA, Meta FAIR, and others); SA-VLA (state-aware action tokenizer improving discrete action reconstruction accuracy); OWMDrive (4D occupancy world model for causally-aware end-to-end driving); WARP (whole-body retargeting from offline human demonstrations, extending mobile manipulation); TacGen (vision-to-tactile alignment and generation to address tactile data scarcity); J-LAW (coupled latent factor graph for joint localization and actionable world modeling); RoboGaze (structured vision-language analysis for evaluating robot world model-generated video); X-Mind (predictive world CoT for Xpeng's end-to-end driving, already reported on the industry side).
Open Source · Tools · Benchmarks
· Ruka-v2: NYU releases a fully open-source tendon-driven dexterous hand covering core degrees of freedom including wrist and finger abduction, aiming to replace "commercial hands costing tens of thousands of dollars" with a low-cost reproducible alternative, lowering the barrier for dexterous manipulation research.
· Qwen-AgentWorld: Alibaba open-sources an agent-oriented "language world model" that lets AI mentally simulate the consequences of an action in language before deciding on the next step.
II. Funding & Deals
Laifual Drive (03952.HK) | Hong Kong IPO | Raised ~HKD 1.07 Billion · hardware
Zhejiang-based Laifual Drive (Chinese harmonic reducer manufacturer) listed on the Hong Kong Stock Exchange on June 30, becoming the first "harmonic reducer stock" on the Hong Kong market, with an IPO price of HKD 85.5 and net proceeds of approximately HKD 1.073 billion, exclusively sponsored by CMB International; funds will be used for capacity expansion and R&D. By 2025 shipment volume, it ranks second among harmonic reducer makers for robots in China with approximately 21.4% market share, and is one of only two Chinese manufacturers to have achieved mass production of harmonic reducers for humanoid robots. 2025 revenue was RMB 261 million (up 142% year-on-year) but with a net loss of approximately RMB 171 million — another example of robot component makers pursuing a "sacrifice margins for volume, rush to capital markets" strategy.Source: Jiazi Guangnian (Chinese tech media) source (WeChat, CN)
Dexmal | Series B | RMB 1 Billion | Valuation Exceeds RMB 10 Billion · embodied
Shenzhen-based embodied AI company Dexmal (Chinese embodied AI startup) has closed a RMB 1 billion Series B, joining the RMB 10 billion valuation club and eyeing an IPO; Shenzhen Capital Group invested in two consecutive rounds, and Lens Technology (founder Zhou Qunfei) upgraded from customer to shareholder. The company's H1 2026 revenue is projected to approach RMB 100 million, with a full-year target of RMB 250–300 million, focusing on foundational world models, physics engines, and humanoid robot deployment — one of the few embodied AI unicorns to put actual revenue on the table.Source: Pencil News (Chinese startup media) source (WeChat, CN)
Liqi Intelligence (Tsinghua-affiliated) | Seed Round | Hundreds of Millions of RMB · world-model
Tsinghua University-affiliated world model startup Liqi Intelligence (Chinese AI startup) has closed a seed round of hundreds of millions of RMB, with Shunwei Capital, Sequoia China, Hillhouse, and Xinglian all participating. The team deliberately downplays the "world model" label, emphasizing coupling data, models, hardware, and infrastructure into systems that actually work in real scenarios — reflecting investors' preference for the "physics + data dual-flywheel" approach.Source: 36Kr source (WeChat, CN)
Yisheng Technology (founded by HKU professor) | Angel Round | Hundreds of Millions of RMB · adjacent
Yisheng Technology (Chinese AI startup), founded by a University of Hong Kong professor, has raised hundreds of millions of RMB in an angel round, focused on "building a memory system for robots" — addressing the long-neglected memory and long-horizon consistency challenges in embodied intelligence.Source: 36Kr Hardware (Chinese tech media) source (WeChat, CN)
NeoWa Robotics | Angel Round | RMB 50 Million · embodied
NeoWa Robotics (Chinese embodied AI startup) has closed a RMB 50 million angel round led by Lanhu Capital with Butong Capital and Gongqingcheng Puyi following; just two months ago it closed a seed round led by Plug and Play China; the founder is the former head of Baidu's autonomous driving and robotics lab, focusing on "embodied intelligence / universal traversal models." Seven other Shanghai companies also closed new rounds on the same day, reflecting continued dense early-stage deal flow in the embodied AI sector.Source: Tech Capital Circle (Chinese media) source (WeChat, CN)
III. Commercial Deployment
Amazon Unveils Next-Generation Warehouse Robot Proteus with Autonomous Package Handoffs · industrial
Amazon has revealed its next-generation autonomous mobile robot Proteus, featuring voice command support and enabling fully autonomous package handoffs from one robot to the next stage in the warehouse. As the world's largest-scale warehouse robotics operator, Amazon's iteration signals a shift in warehouse automation from "point-solution sorting" toward "continuous unmanned material flow."Source: MSN/Fox 59 source
HAI Robotics "Flash Climb" Series Surpasses 10,000 Units in Global Partnerships · industrial
Warehouse robotics company HAI Robotics (Chinese warehouse robot maker) reports its "Flash Climb" series has exceeded 10,000 units in global partnerships, including a collaboration with Arvato to serve a European beauty retailer's omnichannel smart warehouse. Scale has become the moat in box-handling warehouse robotics, though "partnership scale" is a cumulative figure rather than current installed base (⚠️ vendor figures).Source: Ikanchai source
Zhiyuan Robotics' 15,000th G2 Unit Delivered Straight to Factory (Previously Reported) · humanoid
Zhiyuan Robotics' (Chinese humanoid robot company) 15,000th embodied robot Elf G2 rolled off the production line and was delivered the same day to Longqi Technology's factory for production-line quality inspection — less than three months after surpassing 10,000 units in March. This story was already highlighted on June 28; today's follow-up adds supply chain signals including the inauguration of a Tsinghua–Zhiyuan joint research center and the adoption of GaN technology for joint drives at scale.Source: Shangguan News source
IV. Industry Developments
UBTECH Robotics Launches Full-Size Ultra-Bionic Humanoid U1, Priced at RMB 119,800–990,000 · humanoid ⚠️ Pre-sale figures
UBTECH Robotics unveiled its consumer brand "UWORLD" and debut full-size ultra-bionic humanoid robot U1 at its 2026 global launch event, displaying over 50 appearance variants, heights of 1.6–1.85 m, and 88 degrees of freedom across the whole body, targeting home emotional companionship. Pricing ranges from RMB 119,800 for U1 Lite and RMB 169,800 for U1 Pro, to RMB 990,000 for U1 Ultra (male version) and RMB 880,000 (female version); pre-sales launched on JD.com on June 2, with cumulative pre-orders exceeding 11,000 units as of the launch date (pre-sale orders, not deliveries). Founder Zhou Jian stated that robots will replace smartphones as the core AI interaction terminal, with half of the company's energy now directed at home scenarios. This represents the most aggressive attempt yet by China's humanoid sector to expand from "industrial/research" to "consumer companionship," though mass-production delivery and real-world retention remain to be proven.Source: Shanghai Securities News source (WeChat, CN)
Uber and Waymo End Phoenix Robotaxi Partnership, Each Pursues New Direction · autonomy
After nearly three years of operation, Uber and Waymo have ended their Phoenix ride-hailing partnership (the food delivery collaboration had already ended in May 2025). Waymo is reclaiming these vehicles into its own fleet to fulfill autonomous delivery commitments with DoorDash and transit agreements with Via; Uber is pivoting to Lucid + Nuro, planning to deploy more than 20,000 Lucid vehicles equipped with Nuro Driver over six years. Both companies continue their partnership in Austin and Atlanta, and Uber says a new Phoenix partner announcement is coming soon. This marks a shift in the Robotaxi model from "aggregation platforms + third-party fleets" toward vertically integrated control of vehicle capacity.Source: Reuters et al. (×30) source
Apptronik Opens Austin "Robot Park" and Unveils Apollo 2 · humanoid
Apptronik opened its approximately 90,000 sq ft flagship data collection and training facility "Robot Park" on June 30, where fleets of humanoids work on real logistics, manufacturing, and retail tasks to generate training data; it also unveiled humanoid platform Apollo 2 (in bipedal and wheeled variants), featuring 90%+ energy-efficient proprietary actuators, hot-swappable batteries for continuous operation, and expressive interaction. Data will be used to train future humanoid AI models in collaboration with Google DeepMind. The company, which raised $520 million at a $5 billion valuation in February this year, is now materializing its "robot data flywheel" and planning to replicate the Robot Park network across multiple locations.Source: The Star source
YOUIBOT Launches Industrial Embodied Foundation Model FabriX and Native Humanoid "Xifeng" · industrial ⚠️ Vendor figures
YOUIBOT (Chinese industrial robot company) globally launched a new series of embodied intelligence products, including what it calls the "world's first scalable" industrial embodied intelligence foundation model "FabriX (Zhihe)" and the industrial-native humanoid robot "Xifeng," claiming it will empower 10,000 industrial sites within three years. At a time when general-purpose humanoids are crowding into consumer narratives, industrial settings are seen as the most commercially viable near-term deployment ground; however, the "10,000 sites" figure represents cumulative empowerment including integration projects and should be distinguished from actual humanoid unit installations.Source: IItime (Chinese industry media) source (WeChat, CN)
LG Electronics Places Robotics Business Center Under Direct CEO Oversight, Bets on Physical AI · adjacent
LG Electronics has established a Robotics Business Center reporting directly to the CEO, consolidating dispersed operations and accelerating commercialization of physical AI and its home robot "Cloid." Following Samsung and Hyundai, LG is elevating robotics to group-strategy level, consistent with South Korea's broader move to designate physical AI as a national strategic priority.Source: The Korea Times et al. (×4) source
Nvidia Accelerates Hiring for Its China Robotics Team · world-model
Nvidia has posted 12+ new positions for its robotics team in China, interpreted as a talent signal of continued investment in its physical AI and embodied intelligence platforms (GR00T, Cosmos, and others).Source: South China Morning Post source
Hardware · Supply Chain
· STAR ROBOTICS (Chinese robotics startup) launches a 21-DoF fully direct-drive dexterous hand, with motors emerging as a key competitive differentiator — dexterous hands are evolving from mechanical components to highly integrated systems combining micro-motors, precision drives, encoders, drivers, and control algorithms.
· Sonair launches what it claims is the world's first safety-certified 3D ultrasonic sensor (ADAR) for human-robot collaboration near-field collision avoidance, adding a low-cost safety sensing pathway alongside vision and LiDAR (×5 multi-source).
· Dexterous hand valuations are approaching those of Unitree: Lingxin Qiaoshou's (Chinese dexterous hand startup) latest target valuation is approximately $6 billion (around RMB 41 billion), claiming over 80% global market share in high-DoF dexterous hands (⚠️ vendor figures); upstream components including miniature ball screws and harmonic reducers are also heating up simultaneously.
Top comments (0)