Today's Highlights
· Unitree Robotics' STAR Market IPO registration takes effect, targeting RMB 4.202 billion in proceeds — about 73 days from acceptance to registration
· Beijing Humanoid Robot Innovation Center and USTC unveil Labimus, the world's first humanoid manipulation benchmark for precision chemistry labs
· BAAI's Orca world model: pretrained with zero action labels, yet matches dedicated robot systems
· Shanghai Jiao Tong University × Alibaba's LA4VLA: learning language-to-action "blindfolded" first lifts real-robot success rate from 38% to 82%
· JD.com's first full-lifecycle robotics infrastructure project RoboBase breaks ground in Guangzhou as Guangdong province and JD.com sign a comprehensive strategic partnership
I. Research Progress
BAAI's Orca: neither next-frame nor next-action prediction — instead predicting the "next state of the world" · world-model
It's aiming at a more fundamental question: should a world model, like a language model, specialize narrowly in "predicting the next token / next frame / next action"? BAAI's (Beijing Academy of Artificial Intelligence) answer is to build a "world foundation model" — compressing visual and language signals into an abstract internal state, then decoding that state into text, images, or robot actions via pluggable output heads. Training runs on two tracks: an "unconscious learning" track that watches only unlabeled video and predicts the next scene in abstract space, picking up motion, occlusion, and scene dynamics; and a "conscious learning" track that pairs video clips with action descriptions, learning how a given action changes the state. According to the technical report, the base model is pretrained without any action labels, and after attaching an Action Expert trained from scratch, it matches dedicated, specially-trained systems across five robot tasks — if this holds up, it suggests a way around embodied AI's long-standing action-data scarcity. ⚠️ Per technical report
BAAI (Beijing Academy of Artificial Intelligence) · Analysis: The Decoder source
Shanghai Jiao Tong University × Alibaba's LA4VLA: turning vision "off" first, forcing the model to truly learn how language commands action · vla
VLA models have a hidden flaw — target positions and object appearance information in images are so dense that models easily take a "just act on what you see" shortcut, never really learning the language instructions. The authors' diagnostic experiment makes this explicit: swap in conflicting visual scenes, and the model's predicted direction follows the vision instead, showing the language-to-action relationship gets drowned out by strong visual signals. LA4VLA's fix adds a new pretraining stage that temporarily removes visual input, learning continuous action trajectories from language descriptions and robot state alone; to do this, the authors reorganized existing demonstrations into LA-33K (33,000 language-action segments, 1.52 million frames), explicitly extracting the low-level action supervision buried in long trajectories. This step alone lifts MetaWorld's average success rate from 69.7% to 83.0%, while average success across three real-robot tasks (button pressing, book arranging, drink placement) rises from 38.3% to 81.7% — the language patterns learned "blindfolded" transfer back to real visual conditions.
Shanghai Jiao Tong University & Alibaba · Analysis: Embodied Evolution source (WeChat, CN)
Baidu AI Cloud's dVLA-RL: first to connect reinforcement learning to a "discrete diffusion" VLA · vla
Reinforcement learning for VLA models today mostly centers on autoregressive or continuous diffusion policies, with the discrete diffusion approach still lacking a usable RL training paradigm. dVLA-RL, released jointly with Baidu AI Cloud, claims to bridge this gap, enabling discrete diffusion VLA models to also optimize action policies via RL trial-and-error. Specific gains and benchmark details await the full paper; noted here per the release.
Baidu AI Cloud · Analysis: PinWan source
Other papers today: a Social World Models survey — organizing embodied prediction into four tiers, extending from physical world models to world models that "understand human society" (analysis source (WeChat, CN)); LaMem-VLA proposes "latent-native memory" to break VLA's Markovian, moment-to-moment decision-making and improve long-horizon tasks (analysis source (WeChat, CN)).
Open Source · Tools · Benchmarks
· Labimus (Beijing Humanoid Robot Innovation Center × USTC): the world's first humanoid manipulation simulation and benchmark platform for precision chemistry labs. It recreates 30+ pieces of organic chemistry equipment 1:1 with real physical parameters, models powder as individually scoopable rigid particles, and gives real-time balance readings calibrated to a ±0.001g milligram-level tolerance; evaluation moves beyond binary "completed/not completed" to a three-tier progressive scale layered with lighting/texture perturbations, forming a 3×4 evaluation matrix — the first to quantify the precision gap between humanoids merely "being able to" run a chemistry experiment and doing it well. Currently at v0, with liquid handling, multi-step synthesis, and sim-to-real validation planned next. source (WeChat, CN)
· Xingchen Intelligence's (Chinese humanoid startup) StableVLA: released with Peking University and Tsinghua University, aimed at making robot manipulation policies more robust under visual perturbations and reducing failures from visual "confusion."source (WeChat, CN)
· X Square Robot's (Chinese embodied-AI startup) Quanxta Zero: an embodied-AI data collection platform aimed at the industry's widely acknowledged scarcest resource — a deliverable, high-quality data production pipeline. source
· MoXin Technology's (Chinese robotics startup) MoWorld: billed as the first "ultra-high-frame-rate interactive" world model, introducing the Flash World Model concept and emphasizing that high frame rates are essential for real-world usability. ⚠️ Per vendor claim source (WeChat, CN)
II. Funding & Deals
Unitree Robotics | STAR Market IPO registration takes effect | Targeting RMB 4.202 billion · humanoid
The Chinese quadruped and lightweight-humanoid leader Unitree Robotics' STAR Market (Shanghai's Nasdaq-style tech board) IPO registration has officially taken effect, targeting RMB 4.202 billion in proceeds, with reports citing roughly 73 days from acceptance to registration effectiveness. This marks a rare milestone for China's humanoid sector — a move from private equity funding into public capital markets, whereas leading firms have previously stayed unlisted. Proceeds will go toward four areas: expanding core actuator module production lines, iterating humanoid models, industry-specific solutions, and building a nationwide channel network. Unitree is among the few Chinese embodied-AI firms sustaining revenue purely through external market-driven orders (industrial, energy, tourism/culture, university research) without internal group subsidies; however, new production lines still require a 12–24 month build-out, limiting near-term capacity for large-scale procurement, and revenue remains heavily weighted toward hardware unit sales, with a comparatively low share from higher-margin recurring services like maintenance. Source: Star Machine Robotics source (WeChat, CN) | China Securities Journal (Cailianshe) source
dConstruct Technologies (Singapore) | Series A | $125 million · embodied
dConstruct, focused on "reality capture + robotic automation," raised a $125 million Series A — the largest single round among the first cohort of companies incubated by Singapore's RoboNexus accelerator, pointing toward a tech stack that digitizes the physical world with high precision before driving robotic operations on top of it. Source: Embodied Emergence source (WeChat, CN)
Lexiang Technology (Chinese robotics startup) | Pre-A round | Nearly RMB 500 million · embodied ⚠️ Company's own claim
Lexiang Technology, founded just eighteen months ago, closed a Pre-A round of nearly RMB 500 million, led by Ant Group with follow-on investment from Geely Capital and Sanqi Interactive Entertainment. The company states total robot orders have surpassed 30,000 units, with H1 revenue up roughly 600% year-on-year — growth figures and order numbers are the company's own disclosure, pending independent verification. Source: Star Link Robotics source (WeChat, CN)
Chery's Moja Robotics | New funding round | Backed by Agibot, IDG · embodied
Moja Robotics, part of the Chery Automobile group, has drawn investment from Agibot (Chinese humanoid robotics startup), IDG, and others, positioning itself around "automotive-grade" technology applied to commercial service robots — an attempt to transplant automotive supply-chain reliability standards onto robot hardware. Source: Tessy's Embodied Diary source (WeChat, CN)
Yingkan Zhiyi (Chinese bio-inspired flapping-wing robotics startup) | Series A | Tens of millions of RMB · adjacent
Yingkan Zhiyi, a bio-inspired flapping-wing flight robotics company founded by a post-2000 PhD from Shanghai Jiao Tong University, closed a Series A of tens of millions of RMB led by Yuanhe Puhua — its third round in three months, pitched around getting flapping-wing robots to "understand and master fluid dynamics." Source: Yingker source (WeChat, CN)
III. Commercialization & Deployment
JD.com's first full-lifecycle robotics infrastructure project RoboBase breaks ground in Guangzhou as Guangdong signs strategic partnership · adjacent
JD.com signed a comprehensive strategic partnership agreement with Guangdong province, with its first flagship project, RoboBase, breaking ground the same day in Guangzhou — positioned as industry infrastructure covering the full "R&D–testing–deployment–maintenance" lifecycle for robots. Unlike a one-off production-line deployment, this kind of platform is a bet that small and mid-sized robotics manufacturers can share testing and deployment capabilities, lowering the cost of building everything in-house — if it works, it becomes a piece of the embodied-AI industry's basic utility infrastructure. Source: Guangzhou Daily New Express source | Dianshangpai source
Amazon's warehouse robot fleet tops 1 million units, approaching roughly 1.2 million operational jobs · industrial
Per industry analysis, Amazon's global robot fleet has grown from roughly 750,000 units in 2023 to over 1 million (a milestone reached in mid-2025), deployed across 300+ warehouses, with robots involved in about 75% of fulfillment processes; its DeepFleet coordination system reportedly cuts robot travel time by around 10%. Productivity figures tell the trend more clearly: average staffing per warehouse has dropped to about 670 people (a 16-year low), while packages processed per employee rose from about 175 in 2016 to nearly 3,870 in 2025. But Amazon engineers explicitly do not claim robots are "replacing people" — high-touch operations like precisely grasping single items from cluttered bins remain far beyond robotic capability and are an unresolved bottleneck.Source: Silicon Canals source
Agibot–Lens Technology smart robotics production line lands in Huizhou, targeting over 1,000 units of capacity in year one · industrial ⚠️ Per company plan
A smart robotics production line jointly built by Agibot (Chinese humanoid robotics startup) and Lens Technology (Chinese glass/cover-lens manufacturer) has landed in the Huizhou Embodied AI Industrial Park, targeting over 1,000 units of robot production capacity in its first year. This follows Agibot's earlier reported ramp-up to 15,000 units delivered off the production line, but the "1,000 units in year one" figure is a planning target that still awaits production validation. The same day also saw news of the "first Shangrao-made" Agibot robot rolling off the line, with the project reportedly compressed from signing to production within six months. Source: Huizhou Release source (WeChat, CN) | Beautiful Shangrao source (WeChat, CN)
IV. Industry Developments
Waabi's autonomous driving "brain" transfers zero-shot to a Volvo VNL truck · autonomy ⚠️ Per company claim
Autonomous trucking company Waabi says it moved its Waabi Driver — originally trained on the Peterbilt 579 — directly onto a Volvo VNL autonomous truck with entirely different sensors and control systems, without any new data collection, simulation, or fine-tuning, and had it operating autonomously on highways and surface roads from the first mile. If accurate, this kind of zero-shot cross-platform transfer would substantially cut the re-engineering and validation costs typically incurred with each new vehicle platform — a key piece for scaling autonomous trucking. This is currently per the company's own demonstration claims. Source: Fleet Equipment source | Robotics & Automation News
Li Auto "returns to startup mode," doubles down on embodied AI · humanoid
Amid pressure on its core vehicle business, Li Auto is reportedly reorganizing and designating embodied AI as a new growth bet, aiming to extend its intelligent-driving and manufacturing capabilities into robotics. Automakers crossing into embodied AI has become a broader trend (XPeng, GAC, Mitsubishi among others), but there's still a gap between strategic statements and real products. Source: Sina source
Mitsubishi Motors plans humanoid robot mass production in 2027, initially for its own factories · humanoid ⚠️ Per company plan
Mitsubishi Motors is developing humanoid robots jointly with Highlanders, planning mass production in 2027, initially deployed in its own factories to ease Japan's labor shortage. Japanese automakers are increasingly treating humanoid robots as a fallback option for filling production-line labor gaps, but both the production timeline and scale remain at the planning stage. Source: WPS Auto (China auto media) source
Booster Robotics unveils humanoid robot Booster T2 · humanoid ⚠️ Per company claim
Booster Robotics (Chinese humanoid robotics startup) has launched its new humanoid robot, the Booster T2, positioned around flagship-level compute and motion capability for real-world work scenarios; specific specs and mass-production plans are per the release. Source: AI Cloud News source
Hardware · Supply Chain
· Sonair 3D ultrasonic safety sensor: Norway's Sonair has released what it says is the world's first safety-certified 3D ultrasonic sensor, for 3D obstacle avoidance in human-robot collaboration scenarios — pushing collaborative-robot safety detection from 2D light curtains into 3D space. source
· Zhaowei Machinery & Electronics (Chinese micro-motor and gear manufacturer) | Dexterous hand ecosystem: at its new dexterous-hand product launch, Zhaowei Machinery & Electronics signed strategic partnerships with 12 companies including Galbot (Chinese embodied-AI startup), AgiBot's Star (Chinese robotics startup, AstriBot), and Tujian Technology to jointly push dexterous-hand adoption; micro-motor makers Moons' Electric, Leili, and Leadshine are simultaneously moving into this market projected to be worth hundreds of billions of RMB. source (WeChat, CN)
· Humanoid thermal management shifts to liquid cooling: heat dissipation bottlenecks in joint-drive motors and main control chips are becoming more pronounced, with traditional air cooling unable to keep pace with high-power-density motors; liquid cooling is seen as the better path forward, with 6 supply-chain names already named as beneficiaries. source (WeChat, CN)
· Chinese industrial robots' precision shortfall: being able to move boxes isn't the same as handling precision work — Chinese industrial robots still lag international leaders significantly in "precision retention," a real-world constraint on adoption in precision assembly scenarios. source
Top comments (0)