Today's Highlights
· Atoms closes $1.7B in funding led by a16z, with Uber founder Kalanick betting on industrial physical AI
· After a nearly three-month freeze, China restarts robotaxi permits, with Momenta and Baidu the first approved
· Tesla's Q2 operating profit halves, down 57%, as Musk tempers expectations on Optimus mass production
· Psibot (Chinese embodied-AI startup) raises about $100M at a $1.48B valuation, becoming a new unicorn with Chery leading the round
· World models keep attracting capital: HiDream (Chinese world-model startup) closes a RMB 1.5B Series C, with three rounds in three months totaling over RMB 2.1B
1. Research Progress
HOST: A robot learns a new manipulation skill in seconds after watching one human video · manipulation
Existing imitation learning requires a time-consuming training loop for every new skill, and often causes the model to "forget" previously mastered ones. HOST (Human-to-robot One-Shot Skill AcquisiTion) acquires new skills from a single human video in seconds without forgetting existing capabilities, compressing "learn while using" from a training-time loop into a near-instant one-shot alignment — a step toward a continual-learning paradigm closer to how humans learn.
Guangyan Chen et al. · arXiv 2607.20033 source
What actually matters for whole-body humanoid motion tracking? An empirical breakdown · locomotion
Getting a humanoid to follow diverse whole-body reference motions without falling involves numerous modeling and hyperparameter choices whose individual impact is hard to isolate. This paper runs systematic ablation studies on the common factors in recent humanoid motion-imitation pipelines, and open-sources a modular framework, YAHMP, that can be trained, evaluated, and deployed on the Unitree G1 — a directly reusable engineering recipe for teams working on humanoid locomotion control.
Fabio Amadio, Enrico Mingo Hoffman (HUCEBOT) · arXiv 2607.19903 source
Extreme-RGMT: Letting humanoids learn extreme moves while keeping everyday stability · locomotion
General-purpose motion-tracking policies struggle to reliably execute rare, highly dynamic maneuvers, while specialized training often degrades already-learned everyday motions. This paper uses two-stage continual learning to resolve the generalist-versus-specialist trade-off, claiming SOTA on general whole-body motion tracking while significantly improving success rates on extreme moves.
Yubiao Ma et al. · arXiv 2607.20110 source
Toyota Research Institute: Active "drifting" on icy roads turns out to be the optimal collision-avoidance strategy · autonomy
Large-slip drifting has mostly been validated in deliberately staged scenarios, leaving its practical value uncertain. This paper from Toyota Research Institute (TRI) directly answers "when is drifting actually optimal for safety in real winter driving conditions" — on low-friction icy roads, actively entering a high-sideslip drift can, under specific conditions, be the better solution for avoiding a collision, connecting a seemingly flashy maneuver back to serious driving-safety questions.
Elliot Weiss, Thomas Lew, John Subosits et al. (Toyota Research Institute, TRI) · arXiv 2607.19484 source
Milo: The first open-source, low-cost autonomous guide-dog robot · autonomy
Real guide dogs cost roughly $50,000 to acquire and maintain, involve long waitlists, and have short service lives. Milo claims to be the first open-source, low-cost guide-robot platform, performing basic guide-dog collaborative navigation tasks — staying on route, avoiding obstacles and people — fully autonomously onboard with no need for prior environment mapping, advancing accessibility for assisted mobility.
Florian Golemo, Christopher Pal et al. (Mila and others) · arXiv 2607.19530 source
KineBench: A closed-loop checkup for "embodied world models" that bypasses inverse dynamics · benchmark
There has long been no reliable yardstick for evaluating the physical consistency of world models, and nearly all existing closed-loop evaluations rely on an inverse dynamics model (IDM) to extract actions, introducing extra error. KineBench uses explicit kinematic grounding to create the first IDM-free closed-loop benchmark, and finds that embodied video generation exhibits a "nonlinear scaling that plateaus with task complexity" — giving empirical guidance for future data scaling.
Zeyu Liu et al. · arXiv 2607.19876 source
LENS: Having a large model "simplify" cluttered scenes before handing off to the planner · manipulation
Densely cluttered multi-object stacks remain a difficult case for mainstream manipulation methods today — many objects, many collisions, hard-to-predict contact physics. LENS uses a large model to generate a scene-specific, task-specific abstraction layer on the fly, plugging in on top of existing planning/control stacks, and is shown to improve classical planning, model-based control, and a VLA model alike.
Aileen Liao, Rachel Holladay, Dinesh Jayaraman, Michael Posa (UPenn, MIT) · arXiv 2607.19633 source
Other papers today: Koopman Dreamer, using spectrally-constrained deterministic latent dynamics to stabilize long-horizon "imagination" in world models (arXiv 2607.19719 source); Diffusion ReRoll, revisable denoising that improves out-of-distribution performance for robot sequence prediction (arXiv 2607.19919 source); Defer to Plan, adaptive multi-agent fusion for end-to-end V2X driving (arXiv 2607.19774 source); SeededGrasp, using a VLM to predict "seed points" guiding grasping in complex multi-embodiment scenes (arXiv 2607.20207 source); EgoRecovery, learning error-correction ability from first-person human "failure-recovery" videos (arXiv 2607.19745 source); ReferTrack, monocular "refer-then-track" embodied visual tracking (arXiv 2607.20061 source); injecting offline supervision into large-scale VLA reinforcement learning to balance efficiency and out-of-distribution generalization (arXiv 2607.19399 source); NavVerse, an indoor-to-outdoor continuous navigation benchmark (arXiv 2607.19695 source); DINS-IO, self-supervised inertial odometry without position labels (arXiv 2607.20232 source).
Open Source · Tools · Benchmarks
· Unitree UnifoLM-OminiA-0.3: an omni-modal real-time interactive embodied large model, adapted for the mass-production benchmark G1, focused on "single model, full pipeline + direct voice control," the third publicly released node in the UnifoLM family source (WeChat, CN)
· AIRobot (Chinese industrial-VLA startup): with just 0.89B parameters, claims to top an industrial embodied-AI leaderboard and "surpass π0" ⚠️ vendor claim, positioned as a small-parameter end-to-end policy for industrial scenarios source
2. Funding and Deals
Atoms (US) | New round | ~$1.7B · industrial
Led by a16z, with Ben Horowitz joining the board; co-investors include Bain Capital, Fifth Wall, K5 Global, SV Angel, and Uber — the company its founder was forced out of. Bank of America, Goldman Sachs, and JPMorgan are listed as debt partners. Atoms was founded by Uber co-founder Travis Kalanick and focuses on industrial AI for physical sectors such as food, mining, and transportation — using software, sensing, and robotic automation to handle "manufacturing, moving, and warehousing." This is one of the largest single equity raises in physical AI this year, and marks the scale of top-tier US VC bets on industrial-side embodied AI.Source: TechCrunch source
Psibot (Chinese embodied-AI startup) | New round | ~$100M | $1.48B valuation · embodied
Led by automaker Chery, with participation from Lens Technology (a components/sensor supplier to Apple and Tesla) and others; founded in 2024, the company has raised roughly $300M to date and becomes a unicorn with this round. Psibot's focus is end-to-end reinforcement learning embodied models — the Psi R0/R0.5/R1 series — targeting long-horizon task execution, and has already run small-scale sorting validations in a logistics customer's warehouse.Source: Bloomberg, Tech in Asia source
HiDream (Chinese world-model startup) | Series C | RMB 1.5B (over RMB 2.1B across three rounds since April) · world-model
Following an over-RMB-500M round in April and a hundred-million-yuan-class round in May, the company closed another RMB 1.5B Series C in July — three rounds in three months totaling over RMB 2.1B. The funding cadence itself is an industry signal: capital is shifting focus from language models toward "world models," with perception-simulation-interaction of the real world becoming the hottest target. The same day also saw PrimalVerse, a Tsinghua-affiliated team, close a seed round of several hundred million RMB betting on a foundational 4D world model.Source: AI How source (WeChat, CN)
Ropedia (Singapore) | Pre-A | $30M · adjacent
The core team comes from Tsinghua University, Meta, Google, and UC Berkeley, backed by angel investors from Google, Nvidia, and Amazon. Ropedia builds physical AI data infrastructure — proprietary wearable/mobile capture devices that collect, reconstruct, and perceive multimodal human experience in real-world settings, claiming to cut robot-training data collection costs by roughly 50x. Data infrastructure is becoming another hotspot for embodied-AI funding.Source: SiliconANGLE source
LumiBot (Chinese dexterous-hand startup) | Angel round | Tens of millions of RMB · hardware
Led by a fund under Changyou Technology, with two state-owned platform funds co-investing. The founder is a former Wall Street investor; the company focuses on dexterous-hand manipulation solutions and touts what it calls the world's first actively-cooled dexterous hand, Lumi-Dex, saying it has begun sales, with this round's funds going toward mass-production ramp-up.Source: cyzone.com source
Horizon Robotics (9660.HK) | Planned bond issuance | $450M zero-coupon convertible notes · autonomy
To be underwritten by Goldman Sachs and Morgan Stanley, bringing cumulative fundraising to over RMB 20B. As a maker of autonomous-driving chips and solutions, this low-interest convertible bond replenishes capital reserves at a time of deepening cooperation with Volkswagen.Source: Sina Finance source
Previously reported, skipped this round: Weilai Buyuan's near-RMB-1B Pre-A, Agibot's push toward a $20B-valuation IPO, Yuequan Bionic's Pre-A+, and Shihe Robotics' Series C — no material new developments this period (already reported).
3. Commercialization and Deployment
Baidu's Apollo Go begins fully driverless testing in Hong Kong on the 27th · autonomy
In step with the permit thaw in mainland China, Apollo Go is starting fully driverless road testing in Hong Kong, set to formally launch on the 27th — a step toward real driverless operations in a market beyond mainland China. Compared with just holding a road-test permit, this marks a concrete move from "getting licensed" to "on the road."Source: Autohome source
Boston Dynamics' Spot deployed for autonomous inspection at Copper One mine · industrial
Mariana Minerals has deployed Boston Dynamics' quadruped robot Spot at the Copper One copper mine, replacing human workers entering high-risk zones for autonomous inspection — a real production deployment, not a demo. Mine inspection is one of the few scenarios where legged robots have already achieved commercial closure.Source: Mining Weekly source
Fortescue and Liebherr deploy 475 zero-emission autonomous machines at Australian mine · autonomy
Mining giant Fortescue and Liebherr plan to deploy 475 units of zero-emission autonomous machinery at an Australian mine site — a large-scale order for bulk mining automation, pointing to continued scale-up of heavy autonomous vehicles in enclosed industrial settings.Source: Electric & Hybrid Vehicle Technology International source
Renault accelerates warehouse automation, adopts Exotec Skypod sorting system · industrial
Renault has introduced the Skypod warehouse robotics system into its logistics operations, advancing vertical storage and retrieval automation for its spare-parts warehouse — another case of warehouse robotics landing on the automaker supply-chain side.Source: Eureka Magazine source
4. Industry Developments
After a nearly three-month freeze, China restarts robotaxi permits, reopening the driverless-vehicle race · autonomy
According to Bloomberg, Chinese regulators, after completing an industry-wide safety review, resumed issuing robotaxi permits in certain cities starting late June, with Momenta and Baidu among the first approved. Permits had been frozen since April 29, triggered by an incident on March 31 in Wuhan where over a hundred Apollo Go vehicles simultaneously stalled, briefly stranding passengers on a highway. The restart means Baidu, Momenta, and others return to expansion mode, and signals Beijing's intent to keep directly competing with Waymo in autonomous driving. This is the policy signal with the greatest impact on the pace of robotaxi commercialization in China this week.Source: Bloomberg source
Tesla's Q2 profit halves as Musk tempers expectations on Optimus mass production · humanoid
Tesla's Q2 revenue was $28.2B, up 26% year-over-year, with deliveries of 480,000 vehicles setting a Q2 record; but operating profit plunged 57% to about $400M, with operating margin collapsing from 4.1% to 1.4%, and free cash flow turning to a net outflow of $1.1B. As core automotive profitability deteriorates, the market is worried about whether Tesla can sustain its massive investments in Robotaxi, Optimus, and AI infrastructure. Musk called the humanoid robot "the hardest manufacturing ramp in Tesla's history," saying almost no component has an existing supply chain ⚠️ company statement; he also revised the Robotaxi rollout language for Phoenix and Las Vegas to "still in preparation." On Optimus, the company says mass production will begin this year, with the first units delivered to an internal "academy," and part of the Fremont Model S/X line has been converted to Optimus production.Source: Tesla Q2 earnings, Reuters source
Alibaba's Amap unveils "five-in-one" embodied AI stack, touting a unified brain and body · world-model
According to the South China Morning Post, Alibaba's Amap has packaged five components — including perception, mapping, and spatial understanding — into a robotics AI stack, claiming to give robots a "unified brain and body" ⚠️ vendor claim. This is another instance of a major internet company transplanting mapping/spatial capabilities into embodied AI, pointing to positioning around "spatial intelligence as infrastructure."Source: South China Morning Post source
Amazon plans 300-acre robotics manufacturing plant in the US · industrial
Amazon plans to build a 300-acre robotics manufacturing plant in the US, further internalizing warehouse/logistics automation capacity. As one of the world's largest users of warehouse robots, its in-house capacity build-out carries implications for both upstream supply chains and deployment pace.Source: Sohu source
Nvidia PhysX co-founder Zhang Lihua enters physical AI, building world model Fysiverse · world-model
According to an exclusive report, Nvidia PhysX physics-engine co-founder Zhang Lihua has moved into physical AI, leading a team building a next-generation physics-based world model, Fysiverse. A veteran of physics simulation moving into world models reflects how the "differentiable physics + generation" approach is drawing in core talent.Source: NetEase Mobile source
Hardware · Supply Chain
· Leadshine Technology: focused on core robot components, prioritizing joint modules + dexterous hands source
· Jiangsu Leili × MINDSYNC: leveraging three decades of micro-motor expertise to enter dexterous hands, targeting "fingertip stiffness" as a key end-effector pain point source
· Inspire-Robots: humanoid five-finger dexterous hand shipments topped 10,000 units globally in 2025 with over 60% market share; micro servo-cylinders now in production at Xiamen Software Park source (WeChat, CN)
5. This Week's Observations
US weighs banning Chinese humanoid robots, geopolitical rivalry extends into embodied AI · humanoid
According to the South China Morning Post, as US-China tech competition escalates, the US is evaluating import/use restrictions on Chinese humanoid robots. At the same time, Chinese officials have said they will encourage foreign investment into humanoid robotics R&D and related fields. The embodied-AI industry is rapidly being drawn from a pure technology race into supply-chain and market-access geopolitics — worth continued tracking for its impact on overseas expansion paths.Source: South China Morning Post source
MIIT forms data-standards working group; UBTECH leads eight national humanoid-robot standards · humanoid
China's Ministry of Industry and Information Technology (MIIT) formed a TC8 Data Standards Working Group this week and held its first symposium; UBTECH is leading the drafting of eight national standards for humanoid robots, including China's first national standard for evaluating humanoid-robot dataset quality, now formally initiated — signaling a shift in data standards from "existence" to "quality, trustworthiness, and usability." Internationally, the Robot Data Factory (RDF) standard proposed by SK Telecom was adopted by ITU-T for standardization — data standards emerged as a shared focus for regulators this week.Source: Sina Finance source
Washington, D.C. sets clear rules for robotaxis: licensing deposit + per-mile mileage tax · autonomy
Washington, D.C. plans to require a seven-figure-dollar-level deposit for robotaxi operating permits and impose a mileage tax of $0.15 per mile, aimed at preventing empty vehicles from clogging streets and internalizing external costs. This is a concrete example of Western cities shifting from "permitting the technology" to "constraining it with rules," a real cost item for driverless-vehicle unit economics and expansion pace.Source: Hola Chuhai source (WeChat, CN)
Report data: diverging figures on H1 embodied-AI funding in China, unicorn count surges · embodied
According to IT Juzi, total embodied-AI funding in China in H1 2026 exceeded RMB 93.5B across 322 deals, with 26 unicorns valued over RMB 10B (another source cites RMB 104.1B — the figures use different methodologies and scopes and shouldn't be directly compared). GGII released its "H1 2026 China Dexterous Hand Industry Survey Report" over the same period, identifying dexterous hands as the "last inch" of embodied-AI deployment and the segment with the fastest funding and growth; a separate deep-dive report on servo motors estimates demand of 1.2 million sets of joint motors, a market of RMB 91.7B, and a domestic-production rate of 80%. On the cautionary side, the PwC AI Institute warned that "blindly betting on humanoid robots is irrational."Source: AI Intelligence Cube source (WeChat, CN), GGII Humanoid Robotics source (WeChat, CN)
This week's supply-chain roundup: planetary roller screws are the tightest bottleneck, chip localization accelerates · hardware
This week's supply-chain discussion centered heavily on planetary roller screws. These are the core transmission component for humanoid linear joints — a single Optimus unit uses over 30 of them; the bottleneck lies not in design but in manufacturing — constrained by grinding-machine lead times and a shortage of skilled grinding operators, the industry widely reports capacity ramp-up being stuck (one estimate pushes a leading maker's "1,000 units/week" target from July to September), with the domestic-production rate still under 15%. For reference, the global planetary roller screw market was worth only about $1.27B in 2022 — meaning a single customer's annual demand could rival the entire prior global market. A loosening signal: China's first domestic mass-production line came online this week, with Chinese-made units priced at roughly 40-50% of comparable overseas products — supply is being pulled up, but high-end capacity and consistency remain real problems.
A second thread is domestic production of robot-specific chips: at WAIC, several leading players (Agibot, UBTECH, TRON1 and others) have already put Chinese-made MCU/control chips into their robots, with domestic chips moving from "backup" to "standard," and a single-chip full-stack architecture seen as the watershed for scaled deployment (Sunrise's Digua S600 has already locked in 20+ clients); on the cost side, a single humanoid robot often uses 100+ MCUs, with chip BOM cost reaching $3,000-5,000 — a main battleground for domestic substitution. On the battery side, CATL brought online the world's first humanoid-robot battery PACK production line this week, directly tied to humanoid mass production and cost.Source: Neural Leap source (WeChat, CN), AI Melon Field source (WeChat, CN), Humanoid Robot Full-Stack Lobster source (WeChat, CN)
Top comments (0)