Today's Highlights
· The UN's Global Technical Regulation on Automated Driving Systems (co-led by China) was approved at Geneva — the world's first such unified regulation; China's mandatory national standards for L3/L4 have been finalized and are now in the official approval process.
· BMW has launched deployment of Figure 03 humanoid robots for logistics sequencing at its Spartanburg plant, building on an eleven-month Figure 02 pilot in the body shop that assisted in producing over 30,000 X3s.
· Galaxy General Robotics (Chinese humanoid startup) Galbot S1 heavy-load humanoid has entered a CATL battery cell factory, achieving routine production operation in a new-energy manufacturing setting.
· Unitree Robotics (Chinese robot maker) has cut the retail price of its R1 humanoid robot to ¥29,900 with immediate stock availability and no waitlist, marking another step down in China's consumer-grade humanoid price war.
· Commercial service robot leader Pudu Robotics (Chinese service robot company) has closed a new funding round of nearly ¥1 billion, pushing its post-money valuation above ¥10 billion.
I. Research Papers
Causal-rCM: Distilling Autoregressive Video Diffusion into Real-Time Streaming · world-model
Action-conditioned interactive world models have long been bottlenecked by the multi-step sampling cost of video diffusion. This work extends the advanced diffusion distillation framework rCM to autoregressive video diffusion, unifying teacher-forcing and self-forcing training paradigms with an open-source recipe, providing a faster generation path for real-time streaming video and interactive world models. It ranked among the top papers on Hugging Face on the day of release.
Kaiwen Zheng et al. (Tsinghua University) · arXiv 2606.25473 source
In-Context World Modeling: Enabling VLAs to Generalize Across Settings via Contextual World Models · world-model
Most vision-language-action models (VLAs) condition only on the current observation and instruction, failing when the camera viewpoint or robot morphology changes. This method introduces in-context world modeling into the policy, using a small number of interaction demonstrations to infer the dynamics of new settings online, maintaining usability on unseen viewpoints and heterogeneous embodiments — targeting VLAs' widely criticized tendency to fail when the setup changes.
Siyin Wang et al. · arXiv 2606.26025 source
ForceBand: Teaching Robots Contact Force via EMG Signals · manipulation
Human teleoperation is a scalable data source, but motion capture trajectories and internet videos capture only motion and appearance, discarding the contact forces that are critical for force-sensitive manipulation. Jitendra Malik's team uses wearable surface electromyography (sEMG) to simultaneously record muscle effort during action collection, teaching policies "how hard to push," providing a new human-data channel for force-sensitive tasks such as twisting, pressing, and inserting.
Botao He et al. · arXiv 2606.26093 source
Memory Retrieval: General-Purpose Memory Retrieval for Long-Horizon Visuomotor Policies · manipulation
In partially observable environments such as homes, robots need to remember where objects were placed, what step a partner has completed, or when an appliance was switched on. This work proposes a general memory retrieval mechanism that lets visuomotor policies retrieve relevant information from past experience on demand to support long-horizon autonomy, rather than relying on task-specific memory modules.
Rutav Shah et al. (UT Austin) · arXiv 2606.25136 source
Reflective VLA: Incorporating Action Consequences into Context to Improve Generalization · vla
Most VLAs are reactive — predicting the next action from the current instruction and observation alone, implicitly assuming the current observation fully determines the action. This work has the model explicitly incorporate the causal consequences of actions into its context; with this reflection mechanism, VLA generalization and error correction in out-of-distribution scenarios improves.
Qing Lian et al. · arXiv 2606.25215 source
Learning Action Priors: Pre-Training Action Modules for Cross-Embodiment Manipulation · vla
Most VLAs attach an action head to a VLM backbone and train jointly — the vision-language priors are strong, but the action module learns physical motion nearly from scratch. This work learns action priors for the action module separately before cross-embodiment transfer, alleviating data scarcity on the action side, and reports approximately 68% success rate across 13 cross-embodiment tasks.
Dong Jing et al. · arXiv 2606.26095 source · Commentary: Embodied Intelligence Notes source (WeChat, CN)
WOLF-VLA: Extending VLAs to Whole-Body, Contact-Rich Humanoid Motion · locomotion
VLAs have demonstrated strong generalization in robotic arm manipulation, but whole-body, contact-rich humanoid motion remains largely unexplored due to scarce data and a lack of dynamically consistent demonstrations. The DFKI team proposes a VLA framework for humanoid whole-body optimal motion, encoding optimality and safety constraints into the learning pipeline to address this gap.
Melya Boukheddimi et al. (DFKI) · arXiv 2606.25591 source
Other papers today: FORCE (value-calibrated warm-up + self-distillation, improving sample efficiency for VLA reinforcement fine-tuning and breaking the imitation ceiling); GRAFT (part-correspondence-based schematic affordance transfer for few-shot grasping generalization); DeformGen (dynamics-based topological augmentation for deformable object manipulation demonstration expansion); StairMaster (three-stage RL for quadruped climbing high-risk open-riser stairs); RGB (RL-guided whole-body MPPI humanoid control); UniTeD (unified temporal diffusion for joint perception and planning in end-to-end autonomous driving); RAVEN (spatiotemporal memory for long-horizon navigation question answering).
Open Source · Tools · Benchmarks
· TacVerse: A multi-sensor dataset and benchmark for cross-sensor visuotactile perception, evaluating generalization across different tactile sensor designs (arXiv 2606.25877 source).
· 1000 Rallies: An event-camera dataset for robot table tennis plus real-time learned ball state estimation, targeting high-speed, low-latency perception (arXiv 2606.25620 source).
· 3D-Printable Tactile Test Set: Unified 3D-printable texture samples for fair cross-sensor comparison of tactile sensors, improving reproducibility (arXiv 2606.25886 source).
· Realsource: An open-source dataset release for robot training (source: All-About-Industries source).
· Alibaba's Qwen (Chinese AI lab) open-sources an agent-oriented language world model aimed at reducing agent errors; the team claims a 3B-activation model matches leading large models ⚠️ vendor claim (source: Qiling Model Plaza source (WeChat, CN)).
II. Funding & Deals
Pudu Robotics | New Round | ~¥1 Billion | Post-Money Valuation Exceeds ¥10 Billion · adjacent
Longgang Financial Holdings and Asia Investment Capital co-led, with participation from BAIC Industrial Investment, Lens Technology, Shanghai Honghui, and government-guided funds from the Pearl River Delta and Yangtze River Delta regions. Pudu Robotics focuses on commercial service robots (delivery/cleaning) and is extending into embodied AI under a "one brain, multiple forms" approach. According to Frost & Sullivan, it holds a 23% share of the global commercial service robot market, has shipped over 120,000 units, grew revenue more than 100% year-over-year in 2025, and is approaching EBITDA breakeven, with cumulative funding now exceeding ¥2 billion.Source: Hangfan Robotics source (WeChat, CN)
Senad | Series C | ¥300 Million · industrial
Investors include Full Truck Alliance, Haitong Kaiyuan, Oriza Prunus Ventures, Huayi Venture Capital, and Oriza Houwang, with existing investors Dachen Caizhì, Changshi Capital, and Chuhui Capital participating at above-quota allocations. Senad (Chinese logistics robot startup) focuses on loading/unloading and logistics embodied robots, targeting heavy manual tasks such as cargo handling in warehousing and long-haul logistics.Source: Tech Capital Circle source (WeChat, CN)
Luoshi Robotics (Chinese robot arm maker) | Hong Kong IPO (Listing Hearing Passed) · industrial
Luoshi Robotics has passed its Hong Kong Stock Exchange listing hearing, targeting the title of "first full-range intelligent robotics stock" on the Hong Kong market, with CICC and Guotai Junan as joint sponsors. Shareholders include the National Manufacturing Transformation and Upgrade Fund, New Hope Group, and Plum Ventures; its robotic arm business accounts for nearly 70% of revenue.Source: Sina Finance source
Hong Kong Robot IPO Pipeline | Ruiwei Technology, Eacon Mining Technology · autonomy
Concurrently, Ruiwei Technology (Chinese vision-embodied AI firm) has also passed its Hong Kong listing hearing, aiming for the title of "first visual embodied intelligence stock"; Eacon Mining Technology (Chinese autonomous mining truck company, backed by CATL investment and claiming the largest autonomous mining truck fleet globally) has likewise passed its Hong Kong listing hearing. Chinese robotics companies are clustering their Hong Kong listings in mid-2026, with exit channels opening simultaneously across the sector.Source: Sohu source · thebambooworks source
Daxiao Robotics (Chinese embodied AI startup) | H1 2026 Cumulative | Hundreds of Millions of USD · world-model
Daxiao Robotics disclosed cumulative fundraising of hundreds of millions of USD in the first half of 2026. Funds are primarily directed toward its "Kaiwu" world model and integrated hardware-software solutions, emphasizing human-centric ambient data collection to accelerate large-scale embodied training and on-device direct-drive control.Source: Kechuangtong Shares source (WeChat, CN)
III. Commercial Deployment
Galaxy General Robotics Galbot S1 Enters CATL Battery Cell Factory · humanoid
CATL (Chinese battery giant) has selected Galaxy General Robotics (Chinese humanoid startup), with the Galbot S1 heavy-load humanoid entering CATL's battery cell manufacturing lines for routine production operation in a new-energy manufacturing environment. Unlike a one-off demo, "routine operation" means the robot participates in production continuously on standard working hours — making Galaxy General Robotics one of the first humanoid players to enter a heavy-asset manufacturing facility, following leading automakers.Source: Interesting Engineering source
BMW Launches Figure 03 Logistics Sequencing Deployment (Spartanburg) · humanoid
BMW has activated the next-generation humanoid Figure 03 at its Spartanburg, South Carolina plant for complex logistics sequencing tasks: picking parts from unsorted bins, sorting and stacking them onto sequenced carts, and delivering them in order to assembly stations. This builds on an eleven-month Figure 02 pilot at the same plant's body shop — during which it assisted in producing over 30,000 X3s, handled over 90,000 parts, and logged approximately 1,250 work hours. Figure 03 adds a compliant outer shell, wireless charging, voice interaction, and an improved hand with tactile sensors and palm-mounted cameras. Both BMW and Figure AI position this as a milestone validating humanoids' transition "from lab to stable manufacturing workforce."Source: BMW Group source
General Motors Deploys 50 Collaborative Robots, Over 1,000 Positions Affected · industrial
General Motors has deployed approximately 50 collaborative robots at its factories, accompanied by workforce adjustments affecting over 1,000 workers. In contrast to the high media profile of humanoids, the at-scale replacement by mature collaborative robots on production lines more directly reflects the real-world labor impact of automation in manufacturing.Source: Storyboard18 source
Zhiyuan Robotics Spirit G2 Works 6 Consecutive Days on Consumer Electronics Line · humanoid
Multiple Spirit G2 units from Zhiyuan Robotics (Chinese humanoid startup) entered a third-party volume production factory (Longcheer Technology's Nanchang facility) for tablet quality inspection. Starting June 23, the robots worked for 6 consecutive days alongside human workers from 8 a.m. to 7 p.m. with a full livestream; the company claims a task success rate of 99.98%. ⚠️ vendor claim A public livestream is still a validating demonstration rather than a long-term mass production metric; the real test of whether this constitutes a deployable state is full-process compatibility, sustained stability, and low-cost replication — 6 days of data cannot be extrapolated.Source: Jiemian News source
IV. Industry Developments
Unitree R1 Retail Price Cut to ¥29,900 · humanoid
Unitree Robotics has lowered the official starting price of its R1 humanoid robot from ¥39,900 to ¥29,900, with stock available immediately and no waitlist. Following an earlier wave of "price disruptor" narratives at the start of the year, China's consumer- and research-grade humanoid market has pushed further below ¥30,000, shifting the competitive focus from specifications toward scale and cost.Source: 21 Caijing source
Amazon's Zoox Unveils Redesigned Robotaxi · autonomy
Amazon's Zoox has revealed a new-generation steeringwheel-free robotaxi with a boxy form factor and an emphasis on passenger comfort (featuring a panoramic roof), in preparation for upcoming commercial paid operations. This is a key product iteration ahead of its commercial expansion and will compete directly with purpose-built driverless platforms such as Tesla's Cybercab.Source: CNBC source
RoboScience Releases General Embodied Foundation Model Visics · world-model
RoboScience has released a general embodied foundation model, Visics, debuting a VLOA architecture and claiming cross-embodiment, cross-object, and cross-task execution capability at 1/200th the cost of conventional approaches. ⚠️ vendor claim The cross-scenario generalization and cost figures are self-reported from a launch event and lack independent reproduction.Source: Pandaily source
Zhiyuan Robotics Announces "Yuansheng" Ecosystem Plan and Dual World Model Track · world-model
Zhiyuan Robotics has announced dedicated investment of several billion yuan over the next five years to support a developer ecosystem for embodied AI deployment (the "Yuansheng" initiative), and has released five open-source projects. The company splits its world model into two tracks: a "World Action Model" (how actions change the world) and a "World Simulator" (an interactive, inferential, trainable environment). Co-founder Peng Zhihui stated at MWC Shanghai that production volume will exceed 15,000 units, with a new milestone to be disclosed once unit production surpasses 10,000 ⚠️ forward-looking statement.Source: Embodied Intelligence Lab source (WeChat, CN)
Hyundai Workers Plan Strike to Protest Humanoid Robots on the Factory Floor · humanoid
Following Hyundai's full acquisition of Boston Dynamics and plans to deploy Atlas humanoids in its US factories, some Hyundai workers are planning a strike to protest humanoid robots entering their workplace. This marks the first time organized labor has pushed back against humanoid robots in manufacturing, bringing labor relations into a debate that has until now centered on cost and efficiency.Source: Futu NiuNiu source
Hardware · Supply Chain
· Memory: Micron states that a single humanoid robot requires approximately 10× the memory capacity of an L2+ vehicle, and predicts that Optimus could become a larger memory customer than Tesla's cars — humanoid volume ramp-up is seen as a new growth driver for memory demand ⚠️ vendor/analyst claim (source: ifeng.com source).
· Precision Tubing: A company in Ningbo, China manufactures "vascular" tubing for robots, with a minimum wall thickness of 0.005 cm and annual capacity exceeding 1 billion units, entering upstream segments such as joint and liquid cooling circuits (source: Sohu source).
· Power Modules: Midea Welling (Chinese appliance and motor maker) has launched an integrated power module targeting humanoid joints, offering an integrated motor and drive solution aimed at reducing commercialization costs (source: Sina Finance source).
V. Weekly Observations
World's First Autonomous Driving Global Technical Regulation Approved; China's L3/L4 Mandatory National Standard in Final Approval Process · autonomy
At the 199th plenary session of the UN World Forum for Harmonization of Vehicle Regulations (WP.29), held in Geneva from June 22–26, delegates approved the Automated Driving Systems Global Technical Regulation (ADS GTR), jointly led by China, the EU, the UK, the US, Canada, and Japan — the world's first unified regulation of its kind. China served as vice-chair of the GRVA and co-chair of the ADS working group, with the Ministry of Industry and Information Technology (MIIT) organizing the China Automotive Technology and Research Center (CATARC) and the China Academy of Information and Communications Technology (CAICT) to lead the technical drafting. In parallel, MIIT has accelerated work on China's domestic mandatory national standards for automated driving systems; drafting is complete and the standards are now in the official approval process, with more granular safety requirements for L3/L4 and defined safety floors for each level. This provides a unified compliance reference for the autonomous driving industry, and signals that liability determination and market access will shift from self-reported claims to codified standards.Source: Guandian.cn source · MIIT/CATARC announcement
US Moves on Multiple Fronts Against Chinese Robotics: GUARD Act, Commerce Dept. Review, Pentagon Entity List · humanoid
The US government has taken a series of actions against China's robotics sector this month. Bipartisan legislators have introduced the GUARD Act, which would require national security agencies to review humanoid and quadruped robots manufactured in China and other countries. Commerce Secretary Lutnick told business executives on June 22 that the Commerce Department is reviewing imports of robots receiving state subsidies, and that the review could result in executive action steering federal procurement toward domestic manufacturers and restricting certain Chinese robots from sensitive facilities. The Pentagon added Unitree Robotics and other Chinese companies to its list of entities with military ties on June 8; the Department of Defense will be prohibited from renewing direct procurement contracts with those companies from June 30. ⚠️ many measures still at proposal/review stage Most actions beyond the entity list have not yet taken effect, but policy risk on the export and procurement sides has materially escalated.Source: House Select Committee on China · Washington Times source
China Accelerates Humanoid Quality and Component Standards: Maturity Evaluation Framework, Dexterous Hand National Standard · humanoid
Hangzhou Binjiang district has released China's first humanoid robot industry quality management system maturity evaluation standard, drawing on automotive industry experience and covering the full lifecycle from R&D to service. Separately, MIIT's "Humanoid Robotics and Embodied Intelligence Standards System (2026 Edition)" incorporates dexterous hands into the body and component standards, requiring a five-finger structure with at least three degrees of freedom per finger, providing top-level design guidance for industry standardization.Source: Hangzhou Binjiang Market Supervision source (WeChat, CN)
Morgan Stanley Doubles China Humanoid Robot 2026 Shipment Forecast to 50,000 Units · humanoid
Morgan Stanley has raised its forecast for China's 2026 humanoid robot shipments from 28,000 units to 50,000 (up from just 14,000 at the start of the year), citing a faster-than-expected shift from demonstration to commercial deployment. The bank projects annual shipments of 446,000 units by 2030, with market size growing from approximately $2 billion this year to $15 billion, implying a 2025–2030 CAGR of approximately 106%. Catalysts cited include a roughly ¥6.8 billion order from State Grid Corporation of China, policy support, and supply chain capacity expansion. ⚠️ analyst forecast This represents a single firm's projection, not an established fact.Source: Morgan Stanley Research source (WeChat, CN)
Industrial Robot Shipment Forecast: Annual Installations to Exceed 760,000 Units by 2030 · industrial
One industry forecast projects global annual industrial robot shipments will exceed 760,000 units by 2030, driven primarily by AI and automation. Separately, separate data indicate that global robotics funding in 2026 to date has already exceeded ¥60 billion, with capital and production capacity expanding in tandem.Source: Quality Magazine source · Xinji Tanke source (WeChat, CN)
This Week in Supply Chain: Six-Axis Force Sensors See Rapid China-Domestic Substitution; Dexterous Hands and Reducers Remain Bottlenecks · hardware
The most concentrated hardware signal this week is in six-axis force sensors — as humanoids move from "1–2 force-controlled joints" in industrial end-effectors to wrist, ankle, dexterous hand, and eventually whole-body force sensing, demand has structurally stepped up. Multiple white papers and research sources indicate: China's six-axis force sensor market exceeded ¥300 million in 2025, with shipments surpassing 10,000 units and year-over-year growth exceeding 300%; the China-domestic supply ratio has reached approximately 90%, with Changzhou Kunwei (Chinese force sensor maker) holding roughly a 53% share, strain-gauge designs as the mainstream, and MEMS (from companies such as Sensitron) as an emerging alternative. ⚠️ mostly vendor/institutional survey data Scope and methodology vary; absolute figures should be treated with caution. On the actuator side, competition in dexterous hands around lightweighting (PEEK materials) and "high-density motors" is intensifying. Reducers and joint modules are back in focus given reports that Tesla's Optimus V3 is targeting production start at Fremont in July–August with harmonic reducer and joint module sourcing from Taiwan suppliers. Overall, force sensing and transmission are the two segments of the humanoid supply chain under the most strain — and receiving the most capital and policy attention — this week.Source: Xinmaterial Online source (WeChat, CN) · Murong Daxia's Telescope source (WeChat, CN)
Top comments (0)