Today's Highlights
· PokeBot Robotics, founded just 4 months ago, raises a nine-figure Pre-A round led by Shunwei Capital and Matrix Partners China
· UBTECH's consumer humanoid U1 racks up 13,361 pre-orders, first batch to ship September 16
· WeRide partners in Denmark, becoming the first Chinese robotaxi company to enter the Nordics
· Unitree Robotics guides first-half revenue to RMB 1.052–1.128 billion, up over 30% year-on-year
· AgiBot's (Chinese humanoid startup) website now lists 9 partners; Chief Scientist Jean Oh is absent from the list
I. Research Progress
RoboHarness: skip waiting for a universal foundation model — orchestrate VLA, RL, and TAMP instead · manipulation
Long-horizon tasks expose the ceiling of single-policy approaches most clearly: on zero-shot tests on LIBERO-LoHo, roughly four times the length of the original LIBERO, π0.5's full-task success rate is only 6.4%, and even decomposing tasks hierarchically with a world model or LLM before handing off to a VLA tops out at 64.8%. RoboHarness wraps VLA, RL, and TAMP into callable skills, uses a coding agent to route subtasks, and — when adjacent policies have incompatible state distributions — retrieves successful trajectories from the target policy and generates a bridging trajectory via a Memory Bridge to complete the handoff, lifting full-task success to 95.2%. It requires no shared structure, action space, or training data across the underlying policies, and no joint training — this approach shifts the competitive focus from "whose model is stronger" to "how the system schedules."
arXiv 2607.18060 source · Analysis: Jiqizhixin source (WeChat, CN)
CLIFT: turning a closed-source robot foundation model into a humanoid specialist without touching the weights · vla
Most of the strongest robot foundation models are closed-source, leaving users only a "hosted fine-tuning API" — submit data, get back a tuned policy, with no access to weights, gradients, or loss. That locks improvement into pure imitation, ruling out closed-loop methods like reinforcement learning entirely. The authors deliver one of the first empirical studies of this paradigm on a real humanoid robot, using Gemini Robotics On-Device: straightforward API fine-tuning already clearly beats a leading open-source VLA trained on the same demonstrations, but still falls short of deployment-grade performance on agile, contact-rich tasks. CLIFT converts deployment-time reward feedback into API-compatible supervised data, and after two flywheel rounds pushes success rate close to perfect — all without ever "opening the model's box."
Yuxin Chen et al. · arXiv 2607.29172 source
AXIS: can 50,000 crowdsourced web trajectories actually make robot models stronger? · benchmark
No robot to buy, no simulator to set up — teleoperate a Franka Research 3 with keyboard/mouse or a gamepad right in the browser, and each session produces a demonstration trajectory. The question is whether this pause-and-jitter-laden community data can actually train models. Axis Robotics, together with Georgia Tech, UC Berkeley, and Nanyang Technological University, gives a qualified yes: continued pretraining of π0.5 on a fixed snapshot (207 tasks, 50,129 trajectories) lifts LIBERO-Plus overall success rate from 83.9% to 88.8%, rising monotonically as data volume goes from 25%→50%→100%; a RoboCasa365 control group with equal trajectory count and training budget manages only 57.5%. Gains concentrate in sensor noise robustness (+13.7 percentage points) and camera viewpoint changes (+11.3), while lighting and language perturbations see small declines — suggesting what actually matters is task coverage, behavioral diversity, and cleaning/augmentation, not raw trajectory count.
Axis Robotics et al. · arXiv 2607.21588 source · Analysis: Jiqizhixin source (WeChat, CN)
A cross-hardware, cross-policy panorama of legged robot progress, challenges, and opportunities · locomotion
A systematic assessment co-authored by multiple researchers in legged robotics and simulation surveys the current capability boundaries of humanoid and quadruped robots along five axes — hardware, locomotion control, autonomy, data, and applications — and flags the key open problems that must be crossed before large-scale deployment, unusually folding ethics, economic potential, policy implications, and social effects into the outlook as well. At a moment when humanoid discourse is dominated by vendor demos and funding-round tempo, a survey that lays out "what's possible" and "what's still missing" on the same page is worth reading as a reference benchmark.
Jonas Frey et al. · arXiv 2607.28952 source
BWM: an open-source, low-cost, high-fidelity world simulator takes the top spot on the WorldArena leaderboard · world-model
Physical simulators require heavy asset construction and calibration and still leave a sim2real gap, while video generation models lack precise responsiveness to fine-grained robot actions. BWM is an action-conditioned world model that combines initial-environment grounding, dynamic visual history, and temporally aligned action conditioning to perform stateful autoregressive prediction, serving two roles at once — supplementing imitation learning with action-aligned rollout data, and supporting closed-loop policy evaluation, risk anticipation, and policy ranking. Validated on both the WorldArena benchmark and real robots, it ranked first overall across the two application tracks (Track 1 and Track 2) of the WorldArena Challenge; model weights, training and inference code, data-generation, and policy-evaluation interfaces are all open-sourced.
BWM Team · arXiv 2607.29302 source
Mirror Learning: synthesizing "pseudo first-person" expert data from third-person video · manipulation
Humans and animals can learn by watching someone else act, but behavior cloning still only consumes well-aligned first-person data. The authors use a fine-tuned video diffusion model to perform viewpoint transformation, placing the learner into the demonstrator's position, then use an inverse dynamics model to back out an action trajectory in the learner's own control space, synthesizing so-called "mirror data." Experiments show that mirror data alone can train an effective policy, and mixing it into first-person behavior-cloning training further boosts downstream performance — pointing to an alternative path that sidesteps heavy teleoperation data collection.
Yunpeng Liu et al. · arXiv 2607.28737 source
WCM: giving VLA reinforcement learning a critic that predicts the future · vla
Most critic-based reinforcement learning post-training estimates value from a single observation frame or a single VLM hidden layer, which is fundamentally mismatched with the partial observability inherent to robot control. The World Critic Model is built on the lightweight LeJEPA architecture, jointly predicting future latent states and estimating value, so that the critic's representation is explicitly trained to capture temporal dynamics rather than simply regressing a scalar reward.
Senyu Fei et al. · arXiv 2607.29613 source
Auto-JEPA: end-to-end driving doesn't need to reconstruct a complete future world · autonomy
Existing driving world models generally perform dense future prediction — video, occupancy grids, BEV, or other vehicles' motion. The authors argue that planning only needs to care about the scene features that affect the ego vehicle's future actions, so they build the world model as latent-space prediction over continuous intent instead. Results show that future-intent prediction guides the model to focus on planning-relevant visual features, supporting high-quality planning without dense future modeling.
Jiwei Yang et al. · arXiv 2607.29031 source
CorrelationFlow: a training-free method breaks LiDAR scene flow's "monoculture" · perception
The authors argue bluntly that this field has fallen into monoculture — nearly all recent methods share the same feed-forward architecture and the same family of self-supervised losses, inheriting even the same blind spots from each other. CorrelationFlow goes the opposite direction, reducing scene flow to two textbook operations on a BEV occupancy map — connected-component labeling and correlation maximization — entirely training-free, showing that a substantial portion of this problem could have been solved by classical methods all along.
Minh-Quan Dao et al. · arXiv 2607.29237 source
Other papers today: ST-WAM replaces pure pixel-generation supervision with DINOv3 semantic representations to withstand visual distribution shift (arXiv 2607.28993 source); FBFM adds training-free asynchronous feedback to world-action models for online correction during action-chunk execution (arXiv 2607.29235 source); ActFovea gives VLA a plug-and-play runtime safety layer via spatiotemporal visual-action consistency (arXiv 2607.29169 source); Temporal Policy initializes action generation from history to bypass the inference bottleneck of independent Gaussian priors (arXiv 2607.29482 source); TacPrint is a 24-channel capacitive wearable fingertip tactile sensor that can estimate a 35×26 contact depth map (arXiv 2607.29231 source); D-VLC lets a general-purpose VLM directly orchestrate heterogeneous multi-robot collaboration (arXiv 2607.29009 source); Diagnosing Compositional Generalization argues data collection should prioritize covering dependencies in instruction space rather than exhaustively enumerating tasks (arXiv 2607.29687 source); SAGP connects VLM coarse-region reasoning into geometric grasp planning training-free, avoiding grasping knife edges or cup rims (arXiv 2607.29374 source); HAM-VLN uses hierarchical agent memory to cut zero-shot vision-language navigation context by over 65% (arXiv 2607.29600 source); RayViT injects camera geometry into a pretrained ViT so imitation-learning policies withstand viewpoint perturbation (arXiv 2607.29622 source); event-camera upper-body humanoid teleoperation outperforms RGB under high dynamic range and fast motion (arXiv 2607.29227 source).
II. Funding & Deals
PokeBot Robotics ｜ Pre-A round ｜ nine figures (USD) · embodied
The company was founded in April 2026 and closed a tens-of-millions-of-dollars angel round that same month, landing two rounds within four months. This round is co-led by Shunwei Capital and Matrix Partners China, with participation from JQ Capital, Junshan Capital, SEE Fund, Liepin Investment, Yuannuo Capital, Zhongding Capital and other financial and industry investors, plus follow-on from existing shareholders Vision Plus Capital, Xiaomi's strategic investment arm, Honghui Fund, Innoangel Fund, and Oriental Fortune Capital. According to ChinaVenture, the founder previously served as chief scientist at Galbot (Chinese embodied-AI startup). Galbot and BV (Baidu's venture arm) were already shareholders at the angel stage — the mix of industry players and dollar funds moving in together signals that early-stage pricing in the embodied-AI foundation-model race keeps accelerating.Source: Caixin source
Ant Robbyant ｜ first independent funding round ｜ targeting RMB 1.5 billion · world-model ⚠️ Planned figures
According to media reports, Ant Group's embodied-AI unit Robbyant has launched an independent fundraising process, targeting RMB 1.5 billion in its first round with a second round planned by year-end. Robbyant told Securities Times it is "indeed in talks with investors" and will continue focusing on general-purpose robot "brain" models while stepping up investment in embodied-native technology approaches. The company was established in Pudong, Shanghai in March 2025, focused on proprietary embodied-AI foundation models. The financing has not yet closed; amount and timeline are both planned figures.Source: Guandian.cn source
Paxini (tactile sensing) ｜ strategic round ｜ RMB 1 billion · hardware ⚠️ Single-party account
Tactile-sensing and dexterous-hand maker Paxini has secured another strategic round, five months after its last one, raising RMB 1 billion. Investors and valuation were not disclosed.Source: EBRUN source
Agile Robots SE (Germany) ｜ controlling acquisition ｜ India's XNG Automation · industrial
Agile Robots has acquired a majority stake in Bengaluru-based automation firm XNG Automation, bringing the Indian market into its footprint. The same day, WSJ reported the company's revenue is expected to double this year. Deal value undisclosed.Source: TimesTech source
Wanjiyizu ｜ new strategic funding round ｜ RMB 1 billion valuation · adjacent
Equipment-leasing platform Wanjiyizu closed a new strategic round, reaching a post-money valuation of RMB 1 billion. Robot leasing and service platforms continue to draw funding this cycle, complementing the funding heat at the whole-machine level.Source: Dahe Caili Cube source
Zhenlan Instruments completes acquisition of Wuhu Mocap Robotics ｜ closed ｜ RMB 198 million · industrial
Following its late-July disclosure of a plan to acquire 66% of Mocap Robotics for RMB 198 million, Zhenlan Instruments announced the business registration change has been completed, formally closing the deal.Source: Eastmoney source
III. Commercialization & Deployment
WeRide partners with GreenMobility in Denmark, first Chinese robotaxi company to enter the Nordics · autonomy
WeRide has reached a strategic partnership with Danish shared electric mobility platform GreenMobility (NASDAQ Copenhagen: GREENM) to build Denmark's first commercial autonomous shared-mobility project — also making WeRide the first Chinese robotaxi company to establish a presence in the Nordics. Pending regulatory approval, the two plan to launch a public service in the first half of 2027, deploying the L4-level GXR vehicle model compliant with EU regulations. Denmark is WeRide's sixth European country — after announcing Madrid and Zurich in June, the company has added three European markets within just over two months, following an asset-light "platform plus local operator" model: the technology provider supplies the vehicles and autonomous-driving capability, while local partners supply user access channels and operations. GreenMobility operates the Nordic region's largest free-floating car-sharing network, with more than 1,500 electric vehicles. WeRide's business now covers 13 countries and over 40 cities, with an L4 fleet exceeding 3,000 vehicles. The project's rollout still depends on approval progress from the Danish Road Traffic Authority and other regulators.Source: Securities Daily source
UBTECH's consumer humanoid U1 racks up 13,361 pre-orders, first batch delivery starts September 16 · humanoid
The U1 series opened pre-orders in early June and had accumulated 13,361 orders by June 30, with the first batch scheduled for delivery on September 16. The lineup includes three models — Light, Pro, and Ultra — spanning half-body and full-size humanoid form factors, priced from RMB 119,800 to RMB 990,000, marketed around "emotional companionship" — facial and voice interaction and AI conversation — rather than household chores. This marks the first time a consumer humanoid has produced large-scale, real pre-order numbers, but reporting also documents clear divergence in buyer sentiment: a Beijing buyer who ordered a U1 Pro for RMB 159,800 said upfront he doesn't expect it to cook or clean; a buyer in Jiangsu applied for a full deposit refund after watching the online launch event; another buyer who tried it in person in June found the actual movement performance limited. UBTECH founder and CEO Zhou Jian has himself acknowledged the product's capabilities remain clearly limited at this stage, treating emotional interaction as the more viable entry point, with cleaning, laundry, cooking and other capabilities to be layered on later. Real user feedback from September onward will be the true test of these orders.Source: Digital Today, via South China Morning Post source
Yusen Logistics adopts Destro human-robot collaboration platform to overhaul transload operations · industrial
NYK Line's logistics arm Yusen Logistics has partnered with robotics company Destro to deploy an AI-driven human-robot collaboration platform in its transload operations. Deployment scale and site count were not disclosed.Source: Business Wire source
Qatar's unmanned robotaxi trial reports 94% success rate, advances to next phase · autonomy ⚠️ Single-party account
Qatari authorities disclosed a 94% success rate for their unmanned robotaxi trial, on which basis they are advancing to the next testing phase. The methodology, mileage base, and definition of "takeover" behind the figure were not disclosed — it is a self-reported operator metric.Source: Zawya source
Cainiao smart-warehouse projects land in Spain and the Netherlands · industrial
Cainiao (Alibaba's logistics arm) will roll out smart-warehouse projects in Spain and the Netherlands, bringing warehouse automation capability into its European fulfillment network. Project scale and launch timing were not disclosed.Source: Stock Titan source
Tesla expands Robotaxi service to Miami · autonomy
Tesla's Robotaxi service has expanded to Miami, adding another city to its driverless taxi footprint.Source: Chejiahao source
IV. Industry Moves
AgiBot discloses full 9-member partner list for the first time; Chief Scientist Jean Oh absent · embodied
AgiBot Robotics' official website has updated its partner and board lists, and former Chief Scientist Jean Oh's name is no longer among them. AgiBot denied departure rumors to media, saying it was a "role adjustment, not a departure"; however, Oh's personal homepage now lists only an assistant professorship at Shanghai Innovation Institute, and her X profile no longer mentions AgiBot. Oh, a Berkeley PhD, formally joined AgiBot as Chief Scientist in April 2025, leading the formation of its embodied-AI research center overseeing large/small embodied brains, spatial intelligence, world models, and real-robot reinforcement learning, and helped drive AgiBot's collaboration with Physical Intelligence; she was promoted to partner and senior vice president that September. The latest list shows AgiBot has 9 partners in total — Deng Taihua, Peng Zhihui, Jiang Qingsong, Yao Maoqing, Xiong Yan, Wang Chuang, Zhu Jie, Zhang Xiuzheng, and Niu Jia — up from the 7-person list disclosed in September 2025, with notable changes to the rank and responsibilities of figures like Jiang Qingsong and Yao Maoqing. The board comprises 6 executive directors plus 3 independent non-executive directors: Cai Hongjun, Yang Changying, and Zhang Dan. The company launched its Hong Kong IPO process on July 24, so this core-team lineup is being finalized right at that inflection point, amplifying the signal value of the personnel changes.Source: Zhidongxi · Robotics Frontier source
Unitree Robotics updates prospectus: first-half revenue guided to RMB 1.052–1.128 billion, FCC risk disclosed explicitly · humanoid
Following its earlier prospectus response to being included on the FCC's covered list, Unitree Robotics has updated its listing prospectus, disclosing expected January–June 2026 revenue of RMB 1.052–1.128 billion, up 35.62%–45.41% year-on-year. The risk section states explicitly that unless the company obtains an exemption or approval, future robot models it develops face the risk of being unable to be sold in the U.S. market. The company has set August 10 for subscription opening, with listing on the STAR Market possible as early as late August.Source: Guandian.cn source
X Square Robot releases HOST: a robot learns a new task on the spot from a single 29-second human video · embodied ⚠️ Vendor claim
X Square Robot has released the HOST (Human-to-robot One-Shot Skill AcquisiTion) framework, claiming a robot needs only to watch an average 29-second video of an ordinary person performing a task to pick up that skill — completed entirely at inference time, with no model-parameter updates and no teleoperation data collection required. The team tested 50 tasks never seen during training on a dual-arm platform (placing fruit, stacking bowls, wiping plates, inserting pens, folding socks, etc.), all with recorded successes, averaging 62% success rate on the main test set; compared with the strongest SFT baseline, which requires 50 teleoperated demonstrations plus 4 hours of fine-tuning, demonstration volume drops to 1/50 and skill-acquisition time shortens by roughly 507x. Methodologically, the system doesn't imitate human motion directly — it first "imagines" a robot-perspective target image from the human's completed result, then infers the action from that, using dynamic time warping to align progress (progress error dropping from 0.079 to 0.006), with a dual-expert MoT architecture handling vision and action separately. Because parameters are unchanged, old-task performance is barely affected — outperforming the best fine-tuned baseline by 59 percentage points on seven prior tasks (after learning the new task, π0.5+SFT retains only 17% of its original performance, and Wall-OSS+SFT retains 40%). The base model was trained on 193,462 robot trajectories (229 tasks) plus 5,847 paired human-robot videos. All figures above are self-reported by the team.Source: Ifeng Tech source
23 new world-model companies founded in China in seven months, already surpassing all of 2025 · world-model
According to IT Juzi's tally, 23 new world-model companies were founded in China in the first seven months of 2026, compared with 20 for all of 2025 and just 10 in 2024. The density of capital follow-through is even more notable: 18 of these 23 completed their first round or first two rounds of funding within months of founding, 2 were tagged as unicorns at the seed or Pre-A stage, and 4 have entered the RMB-1-billion-plus valuation tier. Sequoia China, Hillhouse Ventures, Tencent Investment, Ant Group, Matrix Partners China, and Shunwei Capital have all moved aggressively, with Xiaomi's strategic investment arm, AgiBot, LimX Dynamics, and Galbot entering as strategic investors. Founder profiles are also diverse — Linguistic Intelligence, founded by former Alibaba Tongyi Qianwen lead Lin Junyang, raised its first round at a post-money valuation of roughly USD 2 billion; Inverse Matrix Technology (incubated by Beijing Academy of Artificial Intelligence), co-founded by 22-year-old Peking University Yuanpei undergraduate Chen Boyuan and PhD student Ji Jiaming, raised a seed round of tens of millions of dollars in March and a follow-on seed++ round exceeding USD 100 million in June. What unites them is writing "world model" into the company's identity from day one, rather than treating it as an appendage to some other business line.Source: 36Kr source
Economic Daily investigation: over RMB 90 billion has flowed into embodied AI in six months, warns against a "future-industry pile-on" · adjacent
Economic Daily published an investigative report on August 3 finding that more than RMB 90 billion flowed into the embodied-AI sector in the first half of this year, with Shanghai, Beijing, Zhejiang and other regions establishing future-industry funds, some reaching RMB 10 billion each in scale. The report flags three tendencies in current regional industrial planning: chasing sectors where the locality has no real advantage, highly similar plans across regions, and clustering into low-value-added segments. According to incomplete statistics, more than 20 of China's 31 provincial-level administrative regions' 15th Five-Year Plans mention developing embodied AI. Qu Shenning, associate research fellow at the Chinese Academy of Social Sciences' Institute of Industrial Economics, said high industry enthusiasm is not itself the problem — the real risk to watch is blind imitation without regard to local fundamentals — citing hydrogen refueling stations as an example: as of end-June 2025, the 560 operating hydrogen stations nationwide served an average of just 43 vehicles per station per day, with most running at a loss.Source: BigGo Finance, via Economic Daily source
Tesla revises Optimus's long-term annual production capacity target from 1 million to 10 million units · humanoid ⚠️ Target figure
Optimus program lead Ashok Elluswamy posted on social media on July 30, revising Optimus's long-term annual production capacity target from 1 million to 10 million units. This is a long-term planning figure, not capacity under construction. On August 3, China A-share humanoid-robot-concept stocks rallied broadly, with Zonergy up for a second consecutive limit-up session and Fulehua hitting its daily limit, as more than 50 concept stocks saw net institutional buying.Source: Guandian.cn source
Fei-Fei Li: the human body is not the optimal solution for any specific task · humanoid
In a recent interview, Fei-Fei Li addressed the humanoid form factor, noting that while the human body's general-purpose design lets it adapt to diverse environments, it is not the optimal solution for any specific task; from a technical-commercialization standpoint, general-purpose structures are precisely the hardest to solve, and in practice specialized form factors are often better suited to specific tasks.Source: The Paper source
Hardware & Supply Chain
· The "Tesla chain" and Leaderdrive: Tesla's Fremont plant has torn out its original Model S/X production line to build an Optimus line, planning annual capacity of 1 million units, with a second line being prepared in Texas; per media reports, Tesla has issued stocking requirements to its supply chain — reaching component supply capacity for 1,000 whole units per week by September this year, ramping to 2,000–2,500 units per week by year-end. Leaderdrive holds over 20% share of China's harmonic reducer market, but in July it pushed back the completion date of its capacity-expansion project ("New-Generation Precision Transmission Device Intelligent Manufacturing Project") to December 2028 — a relatively conservative expansion pace. ⚠️ Stocking requirements and capacity plans are per vendor/media accounts source
· South Korean auto-parts makers pivot to robot components: affected by U.S. FCC restrictions on advanced robots made overseas, South Korean parts makers are accelerating their push into North America. Halla Cast will begin supplying robot semiconductor heat-sink housings to a North American global company in the second half of this year, and has decided to double capacity on its Incheon plant's parts-machining line; SL expects to reach full-scale production in Q3 of leg modules for Boston Dynamics' Spot, while also contract-manufacturing part of Hyundai Motor Group's mobile robot MobED. The South Korean government plans to expand R&D investment and talent-development support for three core component categories: actuators, robotic hands, and sensors source
· Synapticon × Stabilus: the two companies have partnered on robot actuators; scope and product timeline undisclosed source
Top comments (0)