Today's Highlights
· Johnson & Johnson's Ottava surgical robot receives FDA approval, as a medical-device giant formally enters laparoscopic robotics
· Home robotics startup Futuring Robot closes its third funding round in under a year, raising nearly RMB 1 billion in its Pre-A round, becoming the sector's first unicorn
· Fei-Fei Li's World Labs acquires robotics simulation company SceniX, turning generated worlds into training grounds
· US-listed surgical robotics company Vicarious Surgical votes for bankruptcy liquidation, having raised over $2 billion cumulatively
· Fei-Fei Li's team unveils "Masked Visual Actions": a unified world model trained on just 15 hours of data
I. Research Progress
Masked Visual Actions: painting "actions" into pixels, training a unified world model on 15 hours of data · world-model
Video models naturally learn how the world moves, makes contact, and responds — the hard part has always been feeding a robot's "actions" into them in a form they can understand. Stanford's Fei-Fei Li team proposes Masked Visual Actions: representing actions as the "partially revealed trajectory" of some entity in a video — reveal the robot's motion, and the model performs forward dynamics prediction of the scene's response; reveal the target object's desired motion, and the same model can infer the robot action that produces that outcome. Fine-tuned on just 15 hours of real and simulated masked samples, a single set of weights performs policy evaluation, model-based planning, and inverse action synthesis across multiple scenes and embodiments.
Hadi Alzayer, Wenlong Huang et al.（Stanford）· arXiv 2607.19343 source · HF↑3 · Coverage: PaperWeekly source (WeChat, CN)
ergoCub: building "human burden" into hardware optimization, published in Nature Machine Intelligence · locomotion
A team at the Italian Institute of Technology (IIT) has published the humanoid robot ergoCub in Nature Machine Intelligence. Building on iCub3, rather than treating hardware as a fixed chassis, the team uses coupled human-robot dynamics to jointly optimize "ergonomic burden during load-carrying" and "the robot's own locomotion," then translates the optimized structural proportions into a real, fully assembled robot — turning human burden into an optimizable design variable for scenarios like collaborative carrying that reduces worker load.
IIT (Italian Institute of Technology) · Nature Machine Intelligence · Coverage: 科研灵感站 source (WeChat, CN)
Agentic Real2Sim: letting VLM agents automatically bring real-world scenes into runnable simulation · world-model
Real-to-sim (real2sim) has long relied heavily on manual work, since it requires not just reconstructing geometry but also recovering object states, inferring physical parameters, and assembling camera, pose, and trajectory data into a runnable physics simulation. This work automates that pipeline using vision-language agents, pushing toward "generate trainable simulation from a segment of real-world interaction" and lowering the cost of collecting embodied data.
Guanxiong Chen et al. · arXiv 2607.19190 source
WorldScape Policy 2.0: adding "reasoning-augmented memory" to world-action models · world-model
World-action models (WAMs) perform manipulation by jointly modeling visual state transitions and robot actions, but tend to "forget" over long-horizon tasks. Version 2.0 introduces a reasoning-augmented memory mechanism, making world modeling more controllable and steerable over long time horizons and improving consistency across multi-step manipulation.
Haisheng Su et al. · arXiv 2607.18840 source
RoboInter1.5: a full suite of "intermediate representations" for embodied world modeling · benchmark
Existing robotics datasets are expensive, tied to a single embodiment, and lack the fine-grained annotations needed for generalizable reasoning and long-horizon environment dynamics. RoboInter1.5 provides a holistic suite of intermediate representations for embodied world modeling and manipulation, aiming to fill the gap in "reusable, cross-embodiment" structured annotation as a general-purpose interface between world models and policies.
Ziqin Wang et al. · arXiv 2607.18709 source
Athena-Brain: a technical report on an efficient "robot brain" for edge deployment · vla
As embodied agents grow more capable, the industry needs a compact "brain" that can fit onto a robot's own hardware while retaining large models' general knowledge and reasoning. Athena-Brain optimizes for efficiency in edge deployment, aiming to balance general intelligence with high-level interaction with the embodied environment on-device.
Jialian Li et al. · arXiv 2607.18985 source
Other papers today: STeP uses signal temporal logic to give VLAs precise spatial/temporal/logical constraint expression (arXiv 2607.18580 source); MVP-Tac, a miniature visuo-photoelastic dual-mode tactile sensor for minimally invasive surgery (arXiv 2607.18660 source); The Open Ant, an open-source low-cost real-robot reinforcement learning platform simplifying sim-to-real (arXiv 2607.18488 source); No Training, Better Flights uses test-time scaling to improve VLM drone navigation without training (arXiv 2607.19288 source); Recti-Q, out-of-distribution-robust quantization-aware inference for edge robots (arXiv 2607.18540 source); MAGE uses multimodal agents for chip macro-cell placement (arXiv 2607.18536 source).
Open Source · Tools · Benchmarks
· NVIDIA open-sources a medical physics simulation framework: Nvidia released what it calls the "first" GPU-accelerated medical physics simulation framework for training and validating surgical robots, and demonstrated predictive surgical robot capabilities live with CMR Surgical source
· DM0.5: an open-source embodied model that took first place on a simulated manipulation benchmark with a real-robot score of 54.42, advancing alignment between open-source embodied models and real-robot success rates source
II. Funding & Deals
Futuring Robot｜Pre-A round｜nearly RMB 1 billion cumulative · embodied
This round saw continued follow-on investment from existing shareholders Yuanlai Capital, Linkedu Capital, and Boyu Capital, joined by a leading Chinese internet company, CHJ Capital (Inovance's investment arm), Guofang Venture Capital, and Deling Capital, among others. This marks the company's third funding round this year, bringing its cumulative total to nearly RMB 1 billion — a record for the home general-purpose robotics sector. The company focuses on consumer-facing general-purpose home robots and says its products have entered 500+ households. Amid heavy capital inflows into dexterous hands and humanoid platforms, this rare bet on a consumer-facing home robotics company is a footnote to the current shift in robotics funding: renewed enthusiasm, but with sharper focus on orders and real-world deployment.Source: 极客公园 (GeekPark) source (WeChat, CN)
Gritt (United States)｜pre-seed + Series A｜$32.4 million (≈RMB 220 million) · industrial
Of this, a $26 million Series A was led by Obvious Ventures with participation from Union Square Ventures and others. The company was co-founded by Puneet Puri, whose background includes Carnegie Mellon's Robotics Institute, and builds "physical AI" for construction sites: robotic arms plus AI, mounted as attachments onto existing job-site equipment, working alongside construction workers on high-intensity repetitive tasks like material handling and assembly. The company says it has autonomously installed tens of thousands of solar panels with zero damage and has entered job sites for top-ten US construction firms. Its entry point is the labor shortage projected as more than 41% of the US construction workforce is expected to retire by 2031.Source: 智东西 (AI/robotics media) source
Wenshi Robotics｜Series A+｜amount undisclosed · industrial
A provider of full-scenario logistics robotics solutions, this round was backed by Jinfu Capital. Following its Pre-A round in 2022 and Series A in 2023, this marks three consecutive institutional funding rounds, pointing to sustained capital inflow into the AI smart warehousing sector.Source: 第一机器人 (robotics media) source (WeChat, CN)
III. Commercialization & Deployment
Johnson & Johnson's Ottava surgical robot receives FDA approval · adjacent
Johnson & Johnson's Ottava soft-tissue surgical robotic system has received approval from the US FDA. This means a medical-device giant has formally earned entry into the laparoscopic surgical robotics market, long dominated by Intuitive Surgical's da Vinci system — a genuine regulatory milestone rather than a demo or plan, and a substantive signal for the competitive landscape in surgical robotics.Source: MassDevice source
Aurora launches second-generation driverless truck in the US · autonomy
Autonomous freight company Aurora announced the launch of its second-generation driverless truck in the US to meet customer demand. Aurora has already achieved commercial driverless heavy-truck freight operations in Texas; this second-generation launch is an expansion on top of existing real-world operations, not a first-time road debut.Source: Yahoo Finance source
Xinjiang deploys 108-arm cotton topping robot · adjacent
Xinjiang has deployed a 108-arm robot for cotton topping (removing the terminal bud), replacing significant manual labor and reducing labor costs. Open, unstructured agricultural environments have long resisted automation, and this multi-arm, purpose-built form factor is a pragmatic way to push robotics into real-world fields.Source: Big News Network source
Wetour Robotics lands $500,000 warehouse deployment order · industrial
Wetour Robotics has signed a warehouse deployment order worth roughly $500,000, and disclosed that its framework could potentially scale to as many as 20 warehouses across the US. The "20 warehouses" figure is an expansion target, not an already-delivered scale — actual deployment currently stands at this single order.Source: Investing.com
IV. Industry Developments
Fei-Fei Li's World Labs acquires robotics simulation company SceniX · world-model
World-model company World Labs, founded by Fei-Fei Li, has acquired robotics simulation company SceniX (reportedly founded by one of her students). The move directly connects World Labs' generative world models to robotics simulation and training — "using generated worlds as robot training grounds" is a substantive step in world models moving from content generation toward supplying embodied data, echoing this week's broader theme of using video/generated worlds to work around the high cost of real-robot data collection.Source: 具身纪元 (embodied AI media) source (WeChat, CN)
Surgical robotics company Vicarious Surgical votes for bankruptcy liquidation · adjacent
According to foreign media reports, investors in US-listed surgical robotics company Vicarious Surgical (NYSE: RBOT) voted to immediately cease operations and enter bankruptcy liquidation. The company had raised over $2 billion cumulatively. Amid broadly warming robotics funding, a publicly listed surgical robotics company reaching liquidation is a real-world correction to the "story without orders" playbook.Source: 机器人前瞻 (robotics media) source (WeChat, CN)
Volkswagen deepens partnership with Horizon Robotics, L3 autonomous driving to ship as early as H2 2027 · autonomy
Volkswagen Group and Horizon Robotics (Chinese ADAS chipmaker; via joint venture CARIZON) announced a deepened partnership on AI and autonomous driving, with a target of deploying L3-level capability in China as early as the second half of 2027, advancing toward L4 thereafter. This is a further step in binding a Chinese autonomous-driving solution into a multinational automaker's mass-production roadmap.Source: Automotive News source
JD.com acquires Shanghai Zhangjiang land parcel to build 105,000-sqm robotics industrial base · industrial
JD Property has acquired a land parcel in Shanghai's Zhangjiang district, where it will build a robotics-focused industrial base of approximately 105,000 square meters. Combined with recent moves by JD.com, Alibaba, and other major tech companies rushing into robotics infrastructure, this points to large Chinese tech firms shifting resources from pure large-model development toward embodied AI production lines and industrial infrastructure.Source: 观点网 (real estate/industry media) source
Kunlun Tech unveils three world models targeting embodied AI · world-model
Kunlun Tech (Chinese gaming/AI company) unveiled three world models (including the Riemann series) at WAIC, positioned to connect virtual and physical reality and serve embodied AI. The company views world models as a pathway toward larger-scale embodied AI applications; this has been interpreted externally as an "early bet placed ahead of time," with the commercialization path still unproven.Source: 每日经济新闻 (National Business Daily) source (WeChat, CN)
Anthropic reportedly held acquisition talks with embodied-model company Physical Intelligence · embodied ⚠️ unconfirmed report
Multiple foreign media outlets reported that AI company Anthropic held talks to acquire embodied foundation-model company Physical Intelligence (valued at approximately $11 billion), which Physical Intelligence subsequently publicly denied. The report itself reflects intensifying competition among leading digital AI companies for robotic "bodies," though the deal remains unconfirmed.Source: Reuters and others source
Unitree's Wang Xingxing: embodied AI's "ChatGPT moment" could arrive within two to three years · humanoid ⚠️ statement, not milestone
Wang Xingxing, founder of Unitree (Chinese humanoid/quadruped robot maker), said in a public speech that he expects embodied AI's "ChatGPT moment" — a robot completing roughly 80% of tasks in roughly 80% of unfamiliar scenarios — could arrive within two to three years at the earliest. This is an industry outlook rather than a deployment milestone, noted briefly here.Source: Xinhua Finance source (WeChat, CN)
Hardware & Supply Chain
· Solid-state / robot batteries: Reports indicate CATL has brought online what it says is the world's first humanoid robot battery PACK production line, while EVE Energy is also advancing mass-production plans for its 1.0 all-solid-state battery — upstream suppliers are building out dedicated "robot battery" capacity as a standalone product category source (WeChat, CN)
· Servo / joint motors: Industry research estimates demand for humanoid robot joint motors at 1.2 million units, a market size of roughly RMB 91.7 billion, with domestic (Chinese) supply already at roughly 80% — servo motors, as the "core muscle," are becoming a segment of high certainty for Chinese supply chains source (WeChat, CN)
Top comments (0)