Today's Highlights
· onsemi is proposing to acquire edge AI chip maker Synaptics in an all-stock deal worth approximately $7 billion, entering the physical AI space in what would be the company's largest acquisition to date.
· Chinese embodied AI startup Wujie Power (Wujie Dongli) closed an angel round exceeding $200 million led by JD.com, bringing cumulative angel-stage funding to approximately RMB 2.8 billion within six months; the company also claims nearly $100 million in orders.
· General Intuition's $320 million Series A is finalized at a $2.3 billion valuation — led by Khosla Ventures following mid-June reports of negotiations, with participation from Bezos, Schmidt, and others.
· The largest open-source teleoperation dataset to date, ABC-130K (3,500 hours, 130,000 episodes, 195 tasks), has been released along with the full hardware, training, and simulation stack.
· Chinese physical AI foundation model startup Shendu Jizhi has raised another round of several hundred million RMB, focusing on training embodied foundation models using a "human learning" approach.
I. Research
Largest Open-Source Teleoperation Dataset ABC: 3,500 Hours, 130,000 Trajectories · benchmark
A major addition to open-source infrastructure for imitation learning: ABC releases the largest teleoperation dataset to date — ABC-130K — comprising 3,500 hours, over 130,000 episodes, and 195 tasks, together with a low-cost hardware design, training framework, and simulation pipeline, providing a fully reproducible full-stack starting point for behavior cloning.
Arthur Allshire et al. · arXiv 2606.27375 source
World Model "Hallucinations" Are Predictable — and Preventable · world-model
Generative world models are becoming increasingly realistic, yet often silently drift from true dynamics while producing smooth-looking video. Nicklas Hansen and Xiaolong Wang show that such hallucinations are concentrated in low-coverage regions of the state-action space and can be localized and suppressed using lightweight data-side signals, offering a diagnostic approach for reliable world model rollouts.
Nicklas Hansen, Xiaolong Wang (UC San Diego) · arXiv 2606.27326 source
PhysiFormer: Learning Physical Simulation in World Coordinates · world-model
Most video world models predict future frames in viewpoint-dependent pixel space, often producing physically inconsistent results. Andrea Vedaldi et al. propose PhysiFormer, which uses a diffusion transformer to represent objects as 3D meshes in world coordinates: given initial vertex positions, velocities, and material properties (rigid/elastic), it samples future vertex trajectories — a step toward controllable, reusable physical simulation.
Yiming Chen et al. · arXiv 2606.27364 source
Play2Perfect: Dexterous "Play" Pretraining for Precision Assembly · manipulation
Precision assembly is both contact-dense and reward-sparse — difficult for imitation learning to collect data on and for RL to explore. Stanford's C. Karen Liu, Jeannette Bohg, and colleagues systematically analyze what is actually useful in multi-finger "play"-style pretraining, yielding a transferable pretraining recipe for the long-standing challenge of high-precision assembly.
Tyler Ga Wei Lum et al. (Stanford) · arXiv 2606.26428 source
CoStream: Composing Simple Behaviors for Generalizable Complex Manipulation · manipulation
Long-horizon, contact-dense tasks such as inserting a GPU into a PCIe slot require both millimeter-level precision and zero-shot generalization to new tasks. Stanford's Wenlong Huang et al. propose CoStream, which generates complex manipulation by composing simple behaviors, easing the generalization bottleneck of end-to-end policies.
Haonan Chen et al. (Stanford) · arXiv 2606.26423 source
First Generalization Theory for JEPA World Models · world-model
JEPA (Joint Embedding Predictive Architecture), championed by LeCun as a viable path to world models, has demonstrated empirical success without theoretical underpinning. Wang Yisen et al. at Peking University present the first generalization theory for JEPA world models, explaining why predicting dynamics in latent space rather than reconstructing pixels works, providing theoretical grounding for this approach.
Jingyi Cui et al. (Peking University) · arXiv 2606.27014 source
OmniContact: Humanoid Loco-Manipulation via Contact Flow Chaining of Primitive Skills · locomotion
Long-horizon humanoid loco-manipulation requires both robust execution of individual primitive skills and seamless, closed-loop, self-recoverable chaining between them. OmniContact uses a contact flow to chain primitive skills, improving generalization and robustness for humanoid loco-manipulation.
Runyi Yu et al. · arXiv 2606.26201 source
VLA+RL Approach Wins ICRA 2026 LeHome Challenge in Laundry Folding · manipulation
Bimanual laundry folding is a notoriously difficult task given dense contacts and large deformations. This approach won first place in the simulation track among 62 teams at ICRA 2026 LeHome Challenge and runner-up in the real-robot track, by adding a reinforcement learning closed loop to a VLA policy.
Ilia Larchenko · arXiv 2606.27163 source
Other papers today: TaskNPoint — teaching a humanoid to hit a tennis backhand in minutes (2606.26215); "Ace" 8-DOF table tennis robot hardware design targeting professional-level play (2606.26643); Bridging Performance and Generalization — preventing agile flight RL policies from failing upon environment transfer (Scaramuzza group, 2606.27348); Fast LeWorldModel — accelerating JEPA visual planning (2606.26217); Einstein World Models — using counterfactual imagination to extend complex reasoning beyond language (2606.26969); VibeAct — driving contact-dense dexterous manipulation using vibration signals from piezoelectric microphones (2606.27344); HumanoidUMI — driving whole-body humanoid manipulation from robot-free demonstration data (2606.27239).
Open Source · Tools · Benchmarks
· PhysEditWorld: A large-scale dataset for "physically editable" world models that treats physical dynamics as explicit controllable variables rather than implicit correlations (arXiv 2606.26694 source).
· NavIsaacLab: An open-source benchmark for crowd-aware navigation that generates realistic pedestrian crowds in parallel on Isaac Sim (arXiv 2606.26265 source).
· OctoSense: An open-source multimodal sensing platform featuring stereo RGB + event cameras, LiDAR, thermal imaging, IMU, and RTK positioning (UPenn, Kostas Daniilidis et al., arXiv 2606.27317 source).
· ForesightSafety-VLA: A unified safety diagnostic benchmark for VLA models (arXiv 2606.27079 source).
II. Funding & Deals
Wujie Power (Wujie Dongli, Chinese general-purpose embodied AI startup) ｜ Angel Round ｜ Over $200 million (~RMB 1.36 billion) · embodied
Led by a JD.com-affiliated fund, with participation from C Capital, Hony Capital, Shengyu Investment, Fengyuan Investment, and continued backing from existing investors Linear Capital, Sequoia China, Huaye Tiancheng, and Yarui Capital. Founder Zhang Yufeng, formerly head of intelligent driving at Horizon Robotics (Chinese automotive AI chip company), leads the company's general-purpose embodied "large-brain small-brain integrated" approach. The second-generation wheeled dual-arm robot K15 has entered mass production with 1,200 TOPS of compute. Including prior rounds, cumulative angel-stage funding totals approximately RMB 2.8 billion; the company also claims nearly $100 million in global orders (per company disclosure). Source: Caixin source
General Intuition ｜ Series A ｜ $320 million ｜ Valuation $2.3 billion · world-model
Following reports of negotiations in mid-June, the round was led by Khosla Ventures, with participation from Bezos, Schmidt, and others; total disclosed funding reaches approximately $454 million. The company's approach is to train spatiotemporal reasoning capabilities using action labels from gameplay recordings — specifically which buttons were pressed and when — treating large volumes of game replays as training data for AI agents learning to act in the real world. Source: TechCrunch source
Shendu Jizhi (Chinese physical AI foundation model startup) ｜ New Round ｜ Several Hundred Million RMB · world-model
Following a prior round also in the hundreds of millions of RMB, the company has raised another several hundred million RMB to advance physical AI foundation model development. Its approach centers on a "human learning" paradigm, positioning itself against overseas players who bet on human demonstration data, and aiming to train embodied foundation models in a manner more closely modeled on how humans learn. Source: Sina Finance source
Zhicheng AI ｜ Two Rounds in Two Months · world-model
Led by a former CTO of Huawei's U.S. research lab, the company closed two funding rounds within two months. During this period it open-sourced its "Chengling" physical intelligence world model V0.1 and launched updated humanoid products TR4 Pro and TR5 Pro. Source: Zhidongxi source
Oula Wanxiang ｜ Angel Round ｜ Several Hundred Million RMB · embodied
Invested by China Merchants Capital (investment arm of state-owned conglomerate China Merchants Group), with funds directed toward developing an in-house embodied world model, scaling robot hardware manufacturing, and building a global developer ecosystem. Source: Central Enterprise Investment Association source (WeChat, CN)
Juwei Technology ｜ Series B1 ｜ Several Hundred Million RMB ｜ Valuation Nearly RMB 4 Billion · embodied
Closed a Series B1 round in the hundreds of millions of RMB, with a post-money valuation of nearly RMB 4 billion. Source: Guandian.cn source
Qunqing Intelligence ｜ New Round ｜ Nearly RMB 100 Million · industrial
The company focuses on embodied intelligence that "grows from the factory floor," targeting real-world industrial deployment and closed-loop operation. Source: 36Kr Hardtech source (WeChat, CN)
III. Commercialization & Deployment
UBTECH Claims ~5,000 Companion Robot Pre-orders in Two Weeks · humanoid ⚠️ Pre-order figures
UBTECH (Chinese humanoid robot company) reports nearly 5,000 pre-orders for its consumer companion robot (U1) within two weeks, attempting to break into the consumer market. Pre-order numbers remain separate from actual deliveries and retention; whether this converts to sustained shipments remains to be seen. Source: Pandaily source
Waymo Registers in Germany, Continues Expanding Nashville Fleet · autonomy
Waymo has completed legal entity registration in Germany, laying groundwork for European market entry, while continuing to expand its fleet in Nashville, U.S. Registration and road testing remain early-stage steps; large-scale commercial operations are still some distance away. Source: Automotive World source
Cao Cao Mobility Partners with K2 to Deploy First Robotaxis in Abu Dhabi This Year · autonomy ⚠️ Planned figures
Cao Cao Mobility (Chinese ride-hailing platform, Geely subsidiary) has signed a strategic partnership with Abu Dhabi's K2 Group, with plans to deploy its first Robotaxis within the year, continuing China's autonomous driving push into overseas markets. Source: China Daily source
BMW Spartanburg Figure 03 Logistics Deployment (Previously Reported) · humanoid
The Figure 03 deployment at BMW's Spartanburg plant, reported last week, received follow-up coverage this week from autoevolution and other international outlets. No new developments. Source: autoevolution source
IV. Industry Developments
onsemi to Acquire Synaptics for ~$7 Billion, Betting on Physical AI · hardware
U.S. analog chipmaker onsemi has announced an all-stock acquisition of touch and edge AI chip maker Synaptics for approximately $7 billion, at an exchange ratio of 1.350 shares representing a ~19% premium, with the deal expected to close in mid-2027 — onsemi's largest acquisition to date. onsemi aims to extend its capabilities from power and sensing into "intelligent systems," using Synaptics' edge AI compute, human-machine interface, and wireless connectivity to strengthen its physical AI portfolio; the combined company claims it could expand its total addressable market by approximately $30 billion to $243 billion by 2030. onsemi shares fell on the news, with the market skeptical about the valuation and integration outlook. The deal illustrates how competition in physical AI is extending upstream from robot hardware into underlying compute and sensing chips. Source: SiliconANGLE source
UBTECH Launches Wheeled Industrial Humanoid Cruzr Y1 · humanoid ⚠️ Launch/demo figures
UBTECH debuted its wheeled industrial humanoid Cruzr Y1 at an OPPO supply chain technology expo, demonstrating warehouse operations including bin destacking and restacking for factory logistics applications. A trade show demonstration remains distinct from stable mass production and large-scale delivery. Source: Guandian.cn source
Zhiyuan Robotics Releases Embodied Data Collection 2.0, Claims First National Embodied Evaluation Platform · adjacent ⚠️ Vendor claims
Zhiyuan Robotics (Chinese humanoid robot startup) has launched its embodied intelligence "Data Collection 2.0" technology framework and claims to have built China's first embodied intelligence robot evaluation platform, focusing on data acquisition and evaluation infrastructure. Source: Sina Finance source
XPENG Robotics Restructures, He Xiaopeng Takes Direct Charge of Product Division · humanoid
XPENG Robotics has reportedly reorganized into nine divisions, with He Xiaopeng (XPENG CEO) directly heading the product division to accelerate mass production. XPENG also states its second-generation VLA has received global regulatory approval alongside the UN ADS GTR (the regulation itself was previously reported) and plans international launches in 2027. Source: Sina source
LG Electronics and Nvidia Deepen Physical AI Partnership · adjacent
LG Electronics and Nvidia have further deepened their physical AI collaboration to co-build a physical AI ecosystem. The partnership is at a strategic level; specific milestones remain to be seen. Source: Seoul Economic Daily source
Renault Introduces Next-Generation Factory Robot Calvin · industrial
Renault has unveiled its next-generation factory robot Calvin, aimed at automating operations within its manufacturing network. Source: renaultgroup.com source
Hardware · Supply Chain
· Chinese key component localization: Domestic production of three critical components — harmonic drives, ball screws, and six-axis force/torque sensors — is accelerating in China, widely seen in the industry as key to driving down the cost curve for humanoid robots (industry view; specific cost reduction figures not yet confirmed).
· Li Auto develops in-house "Mach M100" chip: Li Auto (Chinese EV maker) has showcased its in-house intelligent driving chip, the Mach M100, a further step toward self-sufficient on-vehicle compute.
Top comments (0)