🤖 AI² Robotics Raises $735M at $3B Valuation for Wheeled Humanoid Robots

• ●Shenzhen-based AI² Robotics closed a $735M round at a $2.8B valuation — one of the largest single raises ever for a humanoid robotics company — backed by state funds, corporate strategics, and financial institutions.
• ●The company’s wheeled humanoid form factor trades bipedal locomotion complexity for mobility and payload efficiency, a deliberate engineering bet that lowers the barrier to near-term industrial deployment.
• ●With this war chest, AI² joins Agility, Apptronik, and Figure in the race to factory-floor scale — but its China-state backing gives it a procurement pipeline that Western rivals cannot easily replicate.
• ●🔎 Read More →
• What matters: A $735M raise at a $3B valuation signals that state-backed Chinese capital is now competing directly with Silicon Valley for humanoid robotics leadership.

🧪 Technology, Research & Innovation

🚗 AUTOPILOT-VQA: A New Benchmark for Testing Vision-Language Models on Safety-Critical Dashcam Incidents

• ●Researchers introduced AUTOPILOT-VQA, a visual question-answering benchmark specifically designed to evaluate whether VLMs and MLLMs can reliably reason about safety-critical driving incidents captured on dashcam footage.
• ●Existing AD benchmarks test scene understanding and trajectory prediction but lack incident-centric evaluation — AUTOPILOT-VQA fills that gap by grounding model assessment in real crash and near-miss scenarios.
• ●The benchmark gives the field a standardized stress-test for VLM safety reasoning, a prerequisite before any such model can be trusted in a production autonomous stack.
• ●🔎 Read More →
• What matters: Without incident-centric benchmarks, VLM safety claims in autonomous driving remain untestable — AUTOPILOT-VQA makes the gap measurable.

🧠 FabriVLA: A Lightweight VLA Model Combines Flow-Matching and Gated Self-Attention for Precise Multi-Task Manipulation

• ●FabriVLA pairs an InternVL3.5 vision-language backbone with a flow-matching action head featuring gated self-attention across action tokens, trained end-to-end from a pretrained VLM via single-stage joint optimization.
• ●The shallow VLM layer fusion strategy enriches spatial context without the inference overhead of deep cross-attention, keeping the model lightweight enough for real-time manipulation control loops.
• ●Precise multi-task manipulation has been a persistent weak point for large VLA models; FabriVLA’s architecture offers a concrete path to closing that gap without scaling to impractical parameter counts.
• ●🔎 Read More →
• What matters: FabriVLA shows that architectural efficiency — not raw scale — may be the decisive variable for deploying VLA models in real manipulation tasks.

🚀 Product, Hardware & Model Launches

🧠 FORT Robotics Extends Physical AI Safety Platform with NVIDIA Halos Integration

• ●FORT Robotics has integrated NVIDIA Halos into its physical AI safety platform, extending hardware-level safety guarantees to AI-driven robotic systems operating in unstructured environments.
• ●NVIDIA Halos provides a safety architecture layer spanning sensors, compute, and actuators — FORT’s integration means its wireless safety controllers can now enforce those guarantees at the edge without cloud dependency.
• ●As physical AI deployments move from controlled pilots to live industrial floors, certified safety stacks like this become a commercial gating requirement, not a differentiator.
• ●🔎 Read More →
• What matters: FORT plus NVIDIA Halos signals that safety certification infrastructure for physical AI is consolidating around a small number of platform partnerships.

🤖 UC San Diego Preclinical Trial: Teleoperated Humanoid Robots Perform Surgery for the First Time

• ●In a UC San Diego preclinical trial, teleoperated humanoid robots successfully completed surgical procedures — the first documented instance of a general-purpose humanoid being used in an operative setting.
• ●Unlike the da Vinci system’s fixed, task-specific arms, humanoid platforms offer reconfigurable dexterity across procedure types, which is the core technical argument for their surgical relevance.
• ●Preclinical success opens the path to IRB-approved human trials, but regulatory timelines for novel surgical robotics platforms typically run three to five years — the clock starts now.
• ●🔎 Read More →
• What matters: The first successful humanoid surgical trial reframes the OR as a near-term deployment environment, not a distant aspiration.

💰 Business, Startups & Investment

🧠 LG Bets Big on Physical AI, Enters World Models Race Against NVIDIA

• ●LG is making a major strategic push into physical AI, directly targeting the world models segment where NVIDIA has staked its Cosmos platform — a rare instance of a consumer electronics giant challenging a compute incumbent on AI infrastructure.
• ●World models — simulation environments that let robots learn physics and causality before real-world deployment — are increasingly seen as the critical moat in physical AI, making LG’s entry a bet on the training data layer, not just hardware.
• ●LG’s manufacturing scale and appliance sensor data give it a credible on-ramp to proprietary world model training sets, which could differentiate its platform from NVIDIA’s more general-purpose Cosmos offering.
• ●🔎 Read More →
• What matters: LG entering the world models race means the physical AI infrastructure layer is now contested by industrial conglomerates, not just AI-native startups.

🚗 AI Startup Tests Level-3 System That Avoids Highway Crashes Without Driver Intervention

• ●An AI startup is actively testing a Level-3 autonomous driving system capable of executing crash-avoidance maneuvers at highway speeds without requiring driver intervention — pushing the boundary of what SAE L3 can handle in safety-critical edge cases.
• ●Level-3 systems legally transfer dynamic driving responsibility to the vehicle under defined conditions; demonstrating reliable crash avoidance without handoff is the hardest technical bar in that envelope and the one regulators scrutinize most closely.
• ●If validated in regulatory testing, this capability could accelerate L3 type-approval in markets like Germany and Japan, where legal frameworks for L3 already exist but certified systems remain scarce.
• ●🔎 Read More →
• What matters: Highway-speed crash avoidance without driver handoff is the single hardest proof point for L3 — a startup claiming it in live testing demands close scrutiny.

⚡China’s Humanoid Capital::AI²’s $735M raise confirms that state-backed Chinese capital is now a primary force shaping the global humanoid robotics competitive landscape.

⚡Safety as Infrastructure::FORT plus NVIDIA Halos shows that physical AI safety certification is consolidating into platform partnerships — a dynamic that will determine which robots get deployed at scale.

⚡VLM Accountability Gap::AUTOPILOT-VQA exposes a critical missing layer in autonomous driving evaluation: no standardized way to test whether AI models reason correctly about crashes.

⚡Humanoids Enter the OR::UC San Diego’s preclinical surgical trial marks the start of a multi-year regulatory clock — the humanoid surgical market is real, but measured in years, not quarters.

⚡World Models as Moat::With LG challenging NVIDIA on world models and AI² flush with capital, the next 18 months will reveal whether physical AI’s decisive advantage lies in simulation infrastructure or in deployment volume — and those may not be the same answer.