Today's Highlights
· AgiBot (Zhiyuan) shipped about 8,400 humanoid robots in H1, overtaking Unitree to claim 44% of the global humanoid market
· Unitree's Shanghai STAR Market IPO subscription was oversubscribed by 8,288x, with an allotment rate as low as 0.0181%
· Archer acquires Boeing's Wisk, Insitu, and SkyGrid; Boeing gets nearly 20% equity in return
· Sony and TSMC invest about $6.4 billion to jointly build a physical AI image sensor production line
· Pony.ai's autonomous driving mileage passes 100 million kilometers; fourth-generation driverless heavy trucks enter mass production
· Knowin (Chinese embodied-AI startup) raises RMB 500 million in an angel++ round; its home robot folds down to under 40 cm
· A Beijing municipal key laboratory for embodied large models, led by Galbot, is inaugurated
· Hidream Driving (Chinese autonomous mining truck maker)'s cumulative driverless mining truck shipments exceed 1,900 units, up 213% year-on-year in H1
I. Research Progress
SimWAM: treat video generation purely as a training signal, and discard the entire video branch at inference · autonomy
World-Action Models bring video dynamics priors to end-to-end autonomous driving, at the cost of having to generate future frames at inference time, keeping latency stubbornly high. SimWAM uses an isolated attention mask so that action prediction depends entirely on nothing but the past — no future frames required — allowing the video branch to be discarded after training, leaving only a self-contained trajectory planner. The video expert and action expert share no parameters and interact only through a unified attention interface, so the video backbone is swappable and the action expert can be scaled independently. It achieves 91.5 PDMS on NAVSIM, with latency significantly lower than prior WAM-based planners, and transfers zero-shot to nuScenes; code and weights are open-sourced. (HF↑22)
Zongchuang Zhao et al. · arXiv 2608.07468 source
DriveTeach-VLA: first teach the model where to look, then teach it how to drive · autonomy
A team from Tsinghua AIR, Beihang University, and Didi argues that autonomous-driving VLA training data skews too heavily toward text-based VQA and chain-of-thought — the language reasoning looks elegant, but attention wanders: a model may say it will "yield to the pedestrian" while its generated trajectory isn't actually focused on that pedestrian. DriveTeach-VLA uses GroundingDINO to label elements like vehicles, pedestrians, and traffic lights for visual self-distillation, then projects expert trajectories from BEV back into image pixel space as prompts, so the base model's existing visual capabilities can be put to direct use. It reaches 90.4 PDMS on NAVSIM, rising to 92.7 with a trajectory selector added; code and models are fully open-sourced.
Yang Yuguang et al. (Beihang University / Tsinghua AIR / Didi) · arXiv 2607.01658 source · accepted at ECCV 2026
AtlasVLA: using only a single wrist camera, yet beating multi-view baselines · vla
The reactive paradigm in VLA has two hard failure modes on partially observable, long-horizon tasks: perceptual forgetting once an object leaves the field of view, and task-progress forgetting across multi-step execution. AtlasVLA proposes a dual-memory architecture — lifting instantaneous 2D observations into a voxel-hashed 4D persistent world-state memory, paired with a working memory that tracks the agent's own history and task progress, both conditioning a diffusion transformer. It improves absolute success rate by 9.4 percentage points on LIBERO-Long and by 17.5 percentage points on real-robot long-horizon tasks.
Guiyu Zhao et al. · arXiv 2608.06729 source
Move the camera and the VLA breaks — this paper fixes it with cross-view action consistency · vla
A VLA fine-tuned on a fixed-camera scene will fail if the camera is simply repositioned, even with the task, objects, instructions, and robot state all unchanged. The authors reset the same LIBERO demonstration to an identical MuJoCo state, render both a nominal and a perturbed viewpoint, and constrain the predicted action-flow velocity fields of the two to agree at the same flow coordinates. This achieves 87.2% on the LIBERO-Plus camera-perturbation track, 7.4 points above pure flow-matching training on the same data and 12.5 points above naive mixed-camera SFT, with no drop in nominal-view performance; on real robots with unseen camera positions, success rate rises from 53.3% to 74.4%. A shuffled-pairing control collapses to 25.8%, confirming the gains come from action-equivalent pairing rather than data volume.
Bingqi Huang et al. · arXiv 2608.06965 source
C2Dex: extracting stable contact from monocular human-hand video, then transferring it to dexterous hands · manipulation
High-quality demonstrations for dexterous manipulation are expensive and scarce, while monocular videos of humans are abundant — the challenge is that reconstructed hand-object contact sequences are temporally unstable and physically implausible, and task-relevant contact geometry gets lost during retargeting. C2Dex aggregates noisy per-frame observations in the object's canonical space to recover stable object-side contact, which then serves both as a trajectory-level constraint for reconstruction and as a transfer target across different hand morphologies. It achieves end-to-end trajectory success rates of 57.78% and 26.67% on DexYCB and TACO, versus 17.78% and 10.00% for the strongest baseline under the same evaluation protocol.
Jie Ren et al. · arXiv 2608.07045 source
A 2.5×3 mm IMU as a tactile sensor, localizing to millimeter precision within 180 ms · perception
In fine manipulation, the fleeting light collisions between a grasped object and its environment are effortless for humans to localize, but robots often lack the resolution and end up making clumsy hard contact. TECDAR embeds a 2.5×3 mm 6D inertial measurement unit in the gripper fingertip, sampling at 7 kHz to capture sub-millisecond deformation while streaming only 84 KB/s of data; fused with robot pose via an extended Kalman filter, it achieves millimeter-level localization within 180 ms, with average precision around 7 mm on line-contact and point-contact tasks — precise enough for millisecond-scale trajectory correction.
Haowen Zheng et al. · arXiv 2608.07075 source
Forward prediction alone isn't enough: grounding the latent space of JEPA world models in physical state · world-model
JEPA-family world models learn action-conditioned latent dynamics from observation sequences, but a pure forward-prediction objective offers no guarantee that individual latents encode the robot's own physical state, or that latent pairs encode state changes — limiting downstream planning and policy learning. PSG-JEPA adds two grounding objectives that apply only during training — tying individual latents to proprioceptive state and latent pairs to multi-horizon joint-angle changes — at zero added cost to the inference architecture or compute. It outperforms prior latent-space world-model baselines across three levels: probe identifiability, goal-conditioned planning in a frozen latent space, and policy learning in both simulation and on real robots.
Haodong Yan et al. · arXiv 2608.06799 source
When collaboration becomes an attack surface: communication attacks on LLM multi-robot systems · benchmark
Prior embodied-safety research has largely focused on single robots, leaving the communication risks of multi-robot collaboration unexamined. The authors construct both an external-entry-point attack setting and an internal-privileged-attack setting, tested across three communication architectures (DMAS, HMAS-1, HMAS-2), three LLMs, and five embodied tasks; unsafe information turns into unsafe actions across all three architectures — DMAS's entry point achieves a 96.7% acceptance rate with 100% activation once accepted, and HMAS-1's unsafe-action success rate reaches 97.8%. Their proposed traceability-and-verification gate (CPV Gate) drops the violation rate from 70.0% to 36.6% — a substantial reduction, but far from a safe margin.
Zhen Huang et al. · arXiv 2608.06830 source
Capek 0.5: organizing embodied VLM capabilities by execution role, not by dataset · vla
Robot execution is iterative — every action reshapes the scene, continually updating what needs to be perceived, reasoned about, and verified. Rather than organizing training by dataset or task, Capek 0.5 splits capabilities into four families by their functional role in the execution chain — spatial reasoning, temporal understanding, action grounding, and state verification — each first learned by a specialist trained with verifiable-reward reinforcement learning from a shared backbone, then merged via weight merging and routing-based policy distillation into a single model. It's released in 2B and 35B-A3B sizes, along with the accompanying Capek-StateBench state-verification benchmark.
Ying Chen et al. · arXiv 2608.06756 source
Other papers today: EMS, an adaptive VLA inference framework, uses a reinforcement-learned switching policy to schedule between a large and a small decoupled system, pushing effective action frequency to 93.4 Hz on LIBERO (arXiv 2608.06434 source); TEMPO splits a VLA's semantic projection layer and action expert into two time scales for RL post-training, so fast updates no longer overwrite high-level semantics (arXiv 2607.07314 source); PILOT has world-action models explicitly model state-transition tokens as a motion chain-of-thought (arXiv 2608.06994 source); WNM-3D adds 3D scene conditioning to world-navigation models for continuous VLN (arXiv 2608.07267 source); Spatiotemporal Agility uses time-constrained RL to let a quadruped perform vision-guided dynamic ball catching (arXiv 2608.06907 source); LyEvO combines Lyapunov stability analysis with statistical model checking for safe sim-to-real transfer (arXiv 2608.06481 source); SoRoMoX implements a differentiable, GPU-parallel Cosserat-rod model for soft robots (arXiv 2608.06650 source); CrossTracer performs cross-embodiment navigation via trajectory residuals (arXiv 2608.06688 source); LifelongCrossNav adds persistent 3D semantic memory to multi-goal cross-floor navigation (arXiv 2608.07079 source); adaptive gait timing gives fault-tolerant locomotion control for large quadrupeds under hardware failure (arXiv 2608.07328 source); Vernata does self-supervised pretraining for LiDAR point representations (arXiv 2608.06919 source); layer-by-layer probing and pruning of planning tokens in driving VLAs, probing how much language-model depth actions actually require (arXiv 2608.07361 source).
Open Source · Tools · Benchmarks
· DRENDS: a robotic endoscope depth dataset for dynamic scenes, published on the Nature journal platform, addressing the shortage of ground-truth depth data in soft-endoscope scenes source
· VicOne Isaac Sim extension: turns robot-attack research presented at DEF CON 34 into a free NVIDIA Isaac Sim extension, enabling robot security testing in simulation source
· Axis Robotics simulation dataset: a new simulation dataset released for robot manipulation research source
II. Funding & Deals
Archer Aviation × Boeing | Acquisition + Strategic Investment | Boeing to hold nearly 20% equity, plus up to $55M additional investment · adjacent
Archer is acquiring autonomous-flight company Wisk Aero, drone maker Insitu, and airspace-management software company SkyGrid from Boeing in one move; both sides signed a definitive agreement on August 10. Consideration is newly issued Archer Class A shares and warrants, giving Boeing roughly nearly 20% equity, along with cross-licensing rights to use Wisk's core autonomous-flight system on Boeing's own civil and defense platforms, plus a commitment to invest up to $55 million in Archer's next funding round. For Archer, this is more than just filling out its autonomous-flight capability: Insitu has more than 3,500 drone systems in service across 35 countries, contributing a profitable defense business with over $200 million in annual revenue, and the nearly 2 million combined flight hours across the three companies will feed into Archer's ZEE AI platform. Archer's stock jumped roughly 20% on the news.Source: Boeing IR / PR Newswire
Sony × TSMC | Joint Venture | About $6.4 billion (¥1 trillion) | Sony ~60% / TSMC ~40% · hardware
The two companies will form a joint venture this year to produce next-generation image sensors at Sony Semiconductor's Kikuyo, Kumamoto Prefecture plant, targeting mass production by 2029 — an investment on the scale of four years' worth of capital expenditure for Sony Group's semiconductor division, with Sony also in talks with the Japanese government over subsidies. The two firms were previously in a foundry-and-designer relationship; shifting to a joint venture is meant to respond faster to AI-driven demand shifts — image sensors are the "eyes" of physical AI, powering robots and self-driving cars. Sony has long held roughly a 50% share of the CMOS image sensor market, but pressure has mounted sharply since Samsung Foundry won the iPhone image-sensor order last August, with both China's OmniVision and Samsung ramping up large investments to catch up.Source: Nikkei, via Seoul Economic Daily source
Knowin (Chinese embodied-AI startup) | Angel++ round | RMB 500 million · embodied
The single round of RMB 500 million was led by Matrix Partners China, with follow-on from Pinpoint Capital, SenseTime Guoxiang Capital, HTI Group, and L2F Guangyuan Founders Fund; Guangyuan Capital served as exclusive financial advisor. Founded in 2025, the company builds consumer-grade embodied-AI large models, with a self-developed GLOW generative embodied-AI model architecture — it first generates physically consistent, structurally 3D first-person data from limited real-robot interaction to fill the data gap, then has a world model predict how an action will change the environment, feeding task outcomes back into a closed loop. The team has grown to over 200 people in under a year. The KNOWIN-X1 home robot, unveiled at this year's World AI Conference, uses a fully foldable dual-arm design that collapses to under 40 cm in height, small enough to sit on a tabletop, sofa, or washing machine; this round's funds will go toward mass-production readiness.Source: PEdaily source
Acorn Robot | Angel round | Amount undisclosed · embodied
The round was co-led by China Merchants Capital and NIO Capital, with follow-on from the Shuimu Tsinghua Alumni Seed Fund, alongside the launch of its "Natus AGE-0 embodied instinct model." This path is a non-consensus bet in an industry currently crowded around VLA: the model uses tactile perception as its core signal source, extracting the underlying rules of physical interaction from contact-mechanics principles; the company claims this enables zero-data cold start and zero-shot generalization across embodiments and materials, forgoing the large-scale data pretraining paradigm.⚠️ Vendor claim Founded in late 2024, the core technical team consists of 9 PhDs and postdocs from Tsinghua's Department of Mechanical Engineering who have worked together on this direction since 2018; commercially, the company has deliberately dropped high-margin custom integration projects to focus solely on standardized dual-arm flexible production cells, selling "manufacturing as a service" for multi-SKU, small-batch scenarios like FMCG, personal care, and food.Source: PEdaily source
Unitree | STAR Market IPO subscription | oversubscribed 8,288.82x | allotment rate 0.0181% · humanoid
Following its August 6 pricing at RMB 150.80/share and an issuance valuation of RMB 60.993 billion, subscription officially opened today. Valid online subscription accounts numbered 9.7846 million, with 53.637 billion valid shares subscribed for — an initial online oversubscription multiple of 8,288.82x, far above the 100x threshold that triggers a reallocation; the lead underwriter shifted 3.236 million shares from the offline tranche to online per the rules, leaving a final allotment rate of just 0.0181%. Each subscription lot of 500 shares requires RMB 75,400 in payment, with RMB 60,000 in Shanghai market holdings needed for a maximum-size subscription; payment for winning allotments is due August 12.Source: Sina Technology / Global Times source
Henggong Precision × AgiBot | Joint venture | Hengzhike Technology established · hardware
Corporate registry data shows Hengzhike Technology is jointly held by Henggong Weichuang Embodied Intelligence Technology, a wholly owned subsidiary of Henggong Precision, and Zhiyuan Innovation Technology, an AgiBot affiliate, with a business scope covering AI foundational resources and technology platforms and intelligent robot R&D. Henggong Precision's core business is castings and precision machining — this kind of "component maker + integrator" joint venture has become increasingly common in the humanoid supply chain.Source: Guandian.cn source
III. Commercialization & Deployment
Pony.ai's global autonomous driving mileage passes 100 million kilometers; fourth-generation L4 driverless heavy trucks enter mass production · autonomy
On August 10, Pony.ai announced that its global autonomous driving mileage has surpassed 100 million kilometers. The more concrete development is on the freight side: its fourth-generation L4-level pure-electric driverless heavy truck has entered mass production, with vehicles rolling off the line over the coming months for deployment on highway trunk routes, Northwest China bulk-commodity dedicated routes, and ports. Its L4-level driverless light truck, launched at the same time, has entered intensive road testing for express delivery, supermarket supply, and cold-chain use. Company vice president He Xing has previously said the moment for large-scale rollout of driverless trucks has arrived. Notably, per the company, cost reductions in the trucking business are mainly driven by supply-chain benefits and cost dilution from thousand-unit-scale Robotaxi production, rather than standalone breakthroughs in truck technology itself. Separately, Pony.ai and JD Auto Service are building a joint autonomous-driving maintenance ecosystem to support Robotaxi fleet operations.Source: Guandian.cn source
Hidream Driving's H1 revenue reaches RMB 804 million; driverless mining truck shipments exceed 1,900 units · autonomy
H1 2026 revenue reached RMB 804 million, up roughly 97% from RMB 408 million in the same period last year; net loss for the period narrowed 75% to RMB 113 million from RMB 455 million a year earlier, with adjusted net loss falling to RMB 19.681 million. The deployment figures say more about where the company stands than the financials do: cumulative shipments of smart-driving products topped 23,000 units as of June 30, with over 3,400 units delivered or pending delivery in the autonomous-driving business, including over 1,900 driverless mining trucks shipped across nearly 40 mines globally, up 213.57% year-on-year in H1. A 500-unit driverless mining truck purchase order signed with Guangna Group in February was the single largest formal procurement order in the global driverless mining truck sector as of end-June; the company also secured over 400 units in letter-of-intent orders for driverless heavy trucks domestically and abroad in H1, and shipped 34 mining robots.Source: Guandian.cn source
FedEx adds dual-arm robots to trailer-loading lines; Amazon to more than double its robotic-arm count this year · industrial
FedEx announced on July 30 that it is deploying Dexterity's dual-arm Mech trailer-loading system at its Hagerstown, Maryland hub, with real-time package-placement decisions made by its Foresight model. Trailer loading has long been one of the hardest links in package logistics to automate — it requires strength, endurance, and real-time adaptation, and FedEx's US network loads tens of thousands of trailers every day. Per its 2026 investor day, the company plans to expand automated loading and unloading to more than 20 US hubs and thousands of dock doors over the coming years. On the same day, Amazon's CFO said on its earnings call that the company plans to more than double its robotic-arm count in 2026, with cumulative robot deployments across its network topping 1 million units since 2012. Industry association A3 has observed that non-automotive industries have recently overtaken automotive as the largest buyers of industrial robots, with Amazon and FedEx driving that shift.Source: Supply Chain Dive source
51WORLD's physical AI core revenue up 545% year-on-year, now 41.4% of total · adjacent
H1 revenue reached about RMB 124 million, up 129.8% year-on-year, with gross margin rising from 41.1% to 44.8% and net loss for the period narrowing 25.5%. Its physical AI core business, 51Sim, saw revenue grow 545.1% year-on-year, with its share of total revenue jumping from 14.8% to 41.4%, turning it from a supplementary business into the main growth driver. The company launched two new products, AperData and AperOne, targeting the twin problems of data scarcity and low fault tolerance in embodied AI, with commercialization beginning in H2; AperOne is targeted for thousand-unit-scale real-world deployment this year.⚠️ Forward-looking claim It also launched development of ECS-1, billed as the world's first commercial remote-sensing satellite dedicated to physical AI, with Huantian Zhihui.Source: Cover News source
Palladyne AI receives additional $2.9M from the US Air Force, bringing total contract value to over $10.6M · embodied
The US Air Force exercised an option on a STRATFI contract, adding $2.9 million and bringing total contract value to over $10.6 million. The new funding is tied to a set of demonstration milestones: mixed-reality teleoperation between a human operator and a remote expert, human-in-the-loop machine learning that lets a robot autonomously learn complex manipulation tasks from a handful of demonstrations, and autonomous execution in Air Force-relevant mission scenarios. The company's Palladyne IQ platform claims to generalize from just 1 to 5 human demonstrations, requiring no cloud connectivity and being hardware-agnostic. The project began in September 2023, with its first phase completed in October 2024.Source: Auganix source
XPeng's GX model line produces its first mass-production Robotaxi · autonomy
The first mass-production Robotaxi based on XPeng's GX model has rolled off the line. Chinese automakers are taking a different path to Robotaxi than fleet operators like Waymo, going straight from their own vehicle production lines to driverless vehicles — making production-line milestones a more meaningful progress marker here than road-test mileage.Source: Autohome source
IV. Industry Developments
AgiBot overtakes Unitree as the world's largest humanoid robot maker, with about 19,100 units shipped globally in H1 · humanoid
According to a report published Sunday by San Francisco-based research firm Smart Analytics Global (SAG), as relayed by the South China Morning Post, Shanghai-based AgiBot (Zhiyuan) shipped about 8,400 humanoid robots from January to June this year, taking 44% of the global market — up 562% year-on-year — surpassing Unitree to claim the top spot globally for the first time; Hangzhou-based Unitree follows with about 5,900 units and a 31% share. The two companies' product mixes differ sharply: AgiBot has spread across multiple lines, including its full-size bipedal A-series, compact X-series, and wheeled G-series, while Unitree's humanoid shipments remain heavily concentrated in its flagship G1, with a substantial share going to education, research, and performance deployments. Zooming out, global humanoid shipments totaled about 19,100 units in H1, up 272% year-on-year, with Chinese makers supplying over 97% of that volume and Chinese buyers accounting for over 85% of demand; the more significant structural shift is that industrial and commercial applications rose from about 50% of the total a year ago to over 70%, overtaking exhibitions and performances as the main driver of volume growth. SAG projects full-year global shipments will approach 60,000 units.⚠️ Compiled estimate The financial gap between the two companies is equally stark: Unitree's IPO prospectus discloses RMB 1.7 billion in revenue and RMB 591 million in adjusted net profit last year; AgiBot has not disclosed financial details, though founder and chairman Deng Taihua said at a partner conference in April that 2025 revenue exceeded RMB 1 billion, with a 2027 target of RMB 10 billion.Source: South China Morning Post source
Beijing Municipal Key Laboratory for Embodied AI Large Models inaugurated, led by Galbot · world-model
On August 9, the Beijing Municipal Key Laboratory for Embodied AI Large Models was inaugurated in Beijing's Haidian District, holding its first academic committee meeting. It is jointly built by Galbot together with Peking University and CATL, marking a provincial/ministerial-level key laboratory recognized in China's embodied-large-model field. The lab's director is Wang He, Galbot's founder and CTO and a Peking University researcher, with the academic committee chaired by Gao Wen, a Chinese Academy of Engineering academician and director of Pengcheng Laboratory, and including 14 Chinese and international academicians and experts such as Zhang Jianwei, Zeng Wenjun, and Zhang Hongjiang, alongside scholars from Peking University, Tsinghua University, Beihang University, and the University of Science and Technology of China. Research will focus on five tracks: fundamental theory for embodied large models, multimodal data infrastructure and standardization, robot operating systems and middleware, and full-machine and key-component development, with three sequential evaluation-cycle goals — an end-to-end embodied large model with full-body and full-hand control within three years, a functionally complete municipal-level key laboratory within six years, and becoming an original-research hub within nine years. At the meeting, Wang He introduced the company's AstraBrain "large brain, small brain" foundation model system, saying that scaling laws have also been validated in motion control.⚠️ Vendor claimSource: China.com source
52 robotics funding deals in China in July; capital is shifting upstream to components · hardware
According to incomplete statistics from Robot Lecture Hall, China's robotics sector recorded 52 funding deals in July 2026 — 23 seed/angel rounds and 17 Series A-stage rounds, with early-stage deals accounting for roughly 76% of the total; by region, Shanghai led with 14 deals, followed by Guangdong with 13 and Beijing with 8.⚠️ Compiled estimate The notable shift is structural: 13 deals, or 25% of the total, went to core-component companies in July, up roughly 5.5 percentage points from about 20% in June — the fastest-growing category that month. Among those 13, tactile/force sensing and joint/transmission/actuator companies tied for the most deals with 5 each, followed by AI chip and data companies with 2. Humanoid and embodied-AI full-machine makers still led overall with 19 deals and a 36.5% share; Robot Era, Yimu Technology, and Weilai Robotics (unrelated to NIO) each closed rounds in the RMB 1 billion-plus range that month — the upstream heat isn't a sign of full-machine makers cooling off, but rather reflects mass-production expectations at the integrator level driving up demand for sensors, joint modules, and dedicated chips.Source: Robot Lecture Hall source
Hardware · Supply Chain
· Dexterous hand and joint-module capacity: Xinuo Future, founded about six months ago, has closed four funding rounds totaling roughly RMB 1.5 billion, planning for 2027 capacity of 50,000 dexterous hands and 50,000 joint modules; Tashan Technology says it holds over 80% share of the humanoid tactile-sensor segment, with orders on hand as of end-May equal to 4x its full-year 2025 revenue ⚠️ Compiled estimate source
· Xinhongye robot cables: on August 10, the company told investors on an interactive platform that its robot cable products do not currently have a supply relationship with Unitree — a reminder that supply-chain "concept stocks" riding Unitree's IPO rally need this kind of claim-by-claim verification source
· Zhongke Huisi's L1/D1/M1 dexterous hands: a joint venture backed by Lens Technology and others has launched three new dexterous-hand products (previously reported) source
Top comments (0)