Today's Highlights
· AgiBot (Chinese humanoid robotics company) rolled off its 15,000th embodied robot, less than three months after its 10,000th unit in late March — claiming a new global record in humanoid mass-production scale and speed. A 6-day factory livestream reported a task success rate of approximately 99.99% (livestream figures).
· Hyundai Motor Group will acquire SoftBank's remaining 9.7% stake in Boston Dynamics for $325 million, achieving full ownership. A production version of Atlas is planned to enter Hyundai's electric vehicle plant in Georgia, USA in 2028.
· Shanghai Innovation Institute and AgiBot released τ0-WM — the world's largest open-source embodied world model, with 5.5B parameters and approximately 27,300 hours of pretraining data, unifying future video prediction, action generation, and action evaluation into a single backbone.
· Tsinghua University open-sourced OpenHLM, a recipe for whole-body loco-manipulation VLA in humanoids: using less than half the demonstration data of strong baselines, it achieves an average task progress of 87.5% on long-horizon tasks (vs. 57.5% for GR00T N1.6).
· A UC San Diego team systematically analyzed world model "hallucinations," categorizing three failure types, showing these can be predicted using label-free signals (ρ≈0.8), and adapting models to new environments using just 50 trajectories — approaching expert-data performance.
I. Research Papers
τ0-WM: World's Largest Open-Source Embodied World Model · world-model
It unifies three previously separate capabilities — predicting future video, generating actions, and evaluating action quality — into a single video DiT backbone, allowing the world model to "imagine" candidate futures before execution and filter out actions that look plausible but fail to advance the task. Pretrained on 5.5B parameters and approximately 27,300 hours of heterogeneous data (including 17,800 hours of real-robot teleoperation), it has been validated on long-horizon dexterous manipulation tasks such as toolbox organization, zipping backpacks, and pipe coupling. Both the model and data are fully open-sourced, making it the largest open-source embodied world model to date.
Shanghai Innovation Institute × AgiBot · arXiv 2606.01027 https://arxiv.org/abs/2606.01027 · Analysis: 真的在读论文 source
OpenHLM: An Open-Source VLA Recipe for Humanoid Whole-Body Loco-Manipulation · locomotion
Addressing the view that "humanoids should be more than bipedal dual-arm platforms," Tsinghua researchers systematically answer how to build whole-body loco-manipulation: collecting data via joint-space whole-body teleoperation, retaining non-humanoid pretraining, generating absolute joint values via multi-step flow, and extending capability with low-cost data sources such as standing teleoperation and the HuMI framework (no real robot required). On the self-built HLM-12 long-horizon task suite, it achieves 87.5% average task progress using less than half the demonstrations of two strong baselines — far exceeding GR00T N1.6 (57.5%) and Ψ0 (48.8%), and approaching the whole-body teleoperation oracle (97.5%).
Tsinghua University Institute for Interdisciplinary Information Sciences (Gao Yang group) et al. · arXiv 2606.22174 https://arxiv.org/abs/2606.22174 · Analysis: 机器之心 source
World Model "Hallucinations" Can Be Predicted — and Fixed with Minimal Data · world-model
Generative world models often render visually smooth sequences that have already diverged from real physics — "silent errors" that are highly dangerous for systems relying on them for planning and control. The research categorizes hallucinations into three failure types: perceptual hallucination, action marginalization, and scene divergence. It then identifies three runtime signals requiring no ground-truth labels (Spearman ρ≈0.8 correlation with real error), arguing that hallucinations are fundamentally a data coverage problem. Using these signals for curiosity-driven data collection, just 50 trajectories suffice to adapt a model to a new environment and approach expert-data performance. The accompanying MMBench2 benchmark covers 427 hours, 210 tasks, and 10 domains.
Nicklas Hansen, Xiaolong Wang (UC San Diego) · Analysis: 具身智能漫谈 source
NavWM: A Unified World Model That Nearly Doubles Navigation Success Rate · autonomy
Most visual navigation policies operate in a "single deterministic future," causing oscillation and looping at complex junctions. NavWM uses a bidirectional Mamba backbone to unify perception, trajectory prediction, and future-frame generation — regressing multiple candidate paths in one pass via trajectory anchors, then using the world model to "visually simulate" each and select the best. At 1.5B parameters, it reduces absolute trajectory error by more than 30% compared to the next-best method, and achieves a zero-shot success rate of 0.44 on unseen scenes — nearly double that of competing approaches.
Institute of Automation, Chinese Academy of Sciences × Beihang University · Analysis: 集智实验室 source
LA4VLA: Decoupling Language-to-Action Pretraining from Vision · vla
In standard VLA training, language signals are overwhelmed by dense visual-action correlations — models appear to follow instructions but actually exploit visual shortcuts. Diagnostic experiments show that when vision and language conflict, models follow vision. Researchers from Shanghai Jiao Tong University and Alibaba address this by explicitly decoupling language-action pretraining: first learning how language instructions constrain continuous actions with no visual input, then combining this with standard VLA training — improving downstream policy performance and robustness under visual perturbation.
Shanghai Jiao Tong University (MINT) × Alibaba · github.com/MINT-SJTU/LA4VLA · Analysis: 具身智能之心 source
Other papers today: A survey on "World Action Models (WAM)" led by Nvidia researchers, mapping the rapidly emerging paradigm of "pretrain to imagine, fine-tune to act"; an ECCV 2026 benchmark of 303 questions revealing significant reasoning gaps in video generation models.
Open Source · Tools · Benchmarks
· Unitree Qmini: Unitree (Chinese robotics company) open-sourced a sub-$1,000 biped robot project with structural parts printable on consumer 3D printers, along with training code — a strategy of trading open-source access for data and ecosystem.
· Alibaba Qwen Language World Model: Alibaba open-sourced an agentic world model covering 7 environments including MCP, search, terminal, and software engineering, claiming to outperform GPT-5.4 on multiple metrics. Note: this is a language world model for agents, distinct from embodied/physical world models ⚠️ vendor figures.
· Zhicheng AI (Chinese physical AI startup) Chengling V0.1: Zhicheng AI open-sourced its "Chengling" physical intelligence world model V0.1 alongside a new funding round, and upgraded its TR4 Pro and TR5 Pro humanoid products.
II. Funding & Deals
Odyssey ｜ Series B ｜ $310 Million ｜ $1.45 Billion Valuation · world-model
Natural Capital led the round, with Amazon, GV, AMD Ventures, EQT, and the CIA's In-Q-Tel participating, along with a preferred cloud partnership with AWS. Founded by autonomous driving veterans, Odyssey focuses on interactive world models and "world simulation." This round makes it a unicorn and is another data point in sustained primary-market appetite for world model companies.
Source: 六观阿尔法 source
RoboScience ｜ Series A ｜ 1 Billion RMB · embodied
This company, founded in late 2024 and previously low-profile, made its public debut at an embodied large-model launch event, disclosing it had closed four funding rounds within a year, with Series A reaching 1 billion RMB. Founder Tian Ye studied under Andrew Ng and previously led Apple's AI platform; Chief Scientist Shao Lin is an NUS assistant professor and two-time ICRA Best Paper winner/nominee. Rather than scaling teleoperation, the company proposes VLOA (adding object trajectory "O" to VLA) combined with a proprietary physics engine, RoboMirage, to build an automated data flywheel — claiming to reduce per-sample data cost from several RMB to fractions of a cent. Cross-embodiment transfer rates and Sim-to-Real performance remain to be validated in real-world settings.
Source: 白鲸实验室 source
SnowOrigin ｜ New Round ｜ Backed by Gong Hongjia, Lu Qi, and Others · adjacent
Gong Hongjia (co-founder of Hikvision, Chinese surveillance giant), Lu Qi (founder of Qizhi Ventures, prominent Chinese tech investor), and overseas institutions have backed this neural interaction company. Using a neural wristband to capture surface electromyographic signals from the forearm, combined with first-person-view devices and AI, the company converts human hand poses and force dynamics into robot training data — targeting the human motion capture layer in the embodied AI data bottleneck.
Source: 高工人形机器人 source
LiberAI ｜ Pre-Series A ｜ Led by Shunwei Capital · world-model
Shunwei Capital (Chinese VC backed by Xiaomi's Lei Jun) led the round, with Cathay Capital, Yuanhe Origin, and Muhua Kechuang participating, and Sequoia China and Zhen Fund continuing to invest. Founded just six months ago with a team of under 30, the company focuses on physical world models and embodied intelligence, pursuing a "human UMI data + world model" approach with accompanying UMI hardware and data collection infrastructure.
Source: 高工人形机器人 source
SEAHI ｜ Series A ｜ Over 1 Billion RMB · adjacent
This round sets a global record for single-round funding in marine robotics. Founded by alumni of Harbin Engineering University, SEAHI specializes in underwater and marine robots — a signal that venture capital is expanding beyond land-based humanoids into underwater and special-purpose applications.
Source: Robot source
III. Commercial Deployment
AgiBot's 15,000th Robot Rolls Off the Line — Deployed Immediately · humanoid ⚠️ Livestream figures
On June 28, AgiBot announced the mass-production rollout of its 15,000th embodied robot (the Elf G2). According to the company, the 10,000th unit (Expedition A3) rolled off on March 30; scaling from 10,000 to 15,000 took under three months, compared to under four months to go from 5,000 to 10,000 — indicating accelerating ramp-up. AgiBot says it has established seven deployment categories: production line loading/unloading, industrial handling, logistics sorting, guided tours, retail, security patrol, and commercial cleaning, spanning full-size, half-size, wheeled humanoid, and quadruped platforms. Units go directly from the line into active deployment. A 6-day livestream connected to the earlier G2 factory broadcast reported approximately 99.99% task success — but this reflects livestream demonstration performance, not mass-production yield. Combined with the "Data Collection 2.0" system released on June 26 (covering 430+ real-world scenarios), the company's logic is to use mass-production scale to feed real-world data back into model and hardware improvement.
Source: 科创板日报 source、AgiBot https://en.prnasia.com/releases/apac/agibot-s-15-000th-robot-rolls-off-the-production-line-marking-a-new-milestone-in-embodied-ai-deployment-538948.shtml
Unitree G1 Powers America's First On-Demand Home Cleaning Service · humanoid ⚠️ Autonomy level unverified
US startup Gatsby has launched the first humanoid robot home cleaning service for general consumers in San Francisco, using Unitree's G1 robot at a flat rate of $150 per visit — below the local human cleaning rate of $150–$300. The company positions itself as a "robot-agnostic" service dispatch layer (Unitree provides the hardware; Gatsby handles scheduling and operations), treating real household task data as an asset to feed back into model improvement. Trial slots are fully booked, but only static images have been released and the proportion of remote human intervention has not been disclosed — leaving the robot's true level of autonomy unclear.
Source: 深观启元 source
Volkswagen ID. Buzz Robotaxi Begins Road Testing in Los Angeles · autonomy
Volkswagen's ID. Buzz autonomous vehicle fleet has begun road testing in Los Angeles in preparation for integration with the Uber platform, marking another move by a legacy automaker into the robotaxi space.
Source: Dimsum Daily https://www.dimsumdaily.hk/vw-id-buzz-robotaxi-begins-los-angeles-trials-ahead-of-uber-launch/
IV. Industry News
Hyundai Motor Acquires SoftBank's Stake for $325 Million, Taking Full Control of Boston Dynamics · humanoid
According to South Korean media, Hyundai Motor Group will acquire SoftBank's remaining 9.7% stake in Boston Dynamics for $325 million, achieving full ownership of the robotics company. SoftBank had retained a minority stake when it sold 80% to Hyundai in 2021 for approximately $880 million, and is now exiting via a contractual put option. Simultaneously, Hyundai plans to sell back to SoftBank its share of the RAI Institute (Robotics and AI Institute) — co-founded with Boston Dynamics in 2022 with cumulative investment of $424 million — for approximately $100 million, effectively splitting assets so that Hyundai retains the hardware company while SoftBank takes the advanced research institute. Hyundai had previously announced plans to deploy a production version of Atlas in its electric vehicle plant in Georgia, USA in 2028; full ownership gives it complete control at the critical juncture of humanoids moving from demonstration to commercial deployment.
Source: 高工人形机器人 source
Faraday Future Unveils Industrial Wheeled-Arm Robot "Faber" · industrial ⚠️ Vendor figures
Faraday Future (FF) unveiled the Faber industrial wheeled-arm robot and a six-series robot lineup at an event in Chicago, claiming entry into industrial automation. Given FF's prolonged track record of delivery delays and cash flow difficulties, the commercial viability of these products will need to be validated by future orders and deployments.
Source: 第一电动网 https://d1ev.com/newsflash/304762
China's First Embodied World Model "Wowu" Completes Generative AI Service Filing · world-model
Beijing Humanoid (National-Local Joint Embodied Intelligence Innovation Center), a state-backed Chinese robotics initiative, has completed China's generative AI service regulatory filing for its embodied world model "Wowu" — described as the first general humanoid robot foundation model to complete the full local compliance review process in China, providing a template for large-model-driven embodied products seeking regulatory clearance in the Chinese market.
Source: 观点网 https://www.guandian.cn/article/20260628/569754.html
Masayoshi Son Delays Retirement, Plans to Transform SoftBank into a "Robot Company" · adjacent ⚠️ Personal statement
Masayoshi Son has said he will delay retirement and continue for more than a decade, aiming to shift SoftBank's focus toward AI and robotics. Combined with SoftBank's move to acquire the RAI Institute from Boston Dynamics, SoftBank's bet on Physical AI is deepening.
onsemi Stock Drops ~24% After Synaptics Acquisition Announcement · hardware
Following its earlier announcement of an all-stock acquisition of edge AI chip maker Synaptics for approximately $7 billion — a move into Physical AI — onsemi shares fell approximately 24%, reflecting market concerns about share dilution and integration risk for what would be the company's largest-ever acquisition.
Source: TIKR.com https://www.tikr.com/blog/on-semiconductor-nasdaq-stock-hits-worst-day-since-march-2024-following-massive-all-stock-acquisition-of-synaptics
Hardware · Supply Chain
· Motors: Nearly 20 Chinese motor manufacturers are competing for humanoid robot joint and actuator orders, making upstream motors one of the most contested segments in the current humanoid supply chain.
· Wiring harnesses: A company in Ningbo, China produces over 1 billion hair-thin robot wiring harnesses annually, supplying the "vascular system" of humanoid robots — continuing the trend of scaled harness assembly seen with players such as Shizhi Zhihang and Tianhai Electronics.
· Reducers / bearings: Research reports have identified automotive component giants such as Schaeffler as key beneficiaries in the humanoid robot supply chain ⚠️ analyst report figures.
· Dexterous hands: AGILINK (spun off from AgiBot), reports cumulative delivery of over 8,000 OmniHand dexterous hands and over 10,000 grippers, claiming operating profitability in its first full quarter of operation ⚠️ unverified single-party figures.
Top comments (0)