Deepwise raised hundreds of millions across two funding rounds; physical AI market accelerates.

Physical AI as the next generation of artificial intelligence represents a core strategic direction and contested frontier in global technology competition. The industry currently faces critical bottlenecks: the absence of general-purpose foundation models, high implementation costs, and weak generalization capabilities. Building domestically autonomous and controllable physics AI foundation models has become an essential requirement for industrial advancement.

As a flagship company in the physics AI sector, Deep Insight pioneered the "human learning" original technology roadmap and, leveraging an industry-leading full-stack R&D team, achieved complete commercialization across data, models, ontology, and applications within just one year of founding. Its core technology development is nearly a year ahead of comparable U.S. projects, providing a viable solution to the industry's shared pain point of difficult implementation. All product lines currently generate commercial orders, with cumulative contract value exceeding tens of millions of yuan, establishing it as one of the fastest-landing full-stack technology companies in the sector.

Deep Insight recently announced a new financing round worth hundreds of millions of yuan. The round is led by China Life Long Triangle S&T Innovation Fund, with continued investment from existing shareholders Puhua Capital and Chengtong S&T Innovation Fund, alongside market and industrial capital participation from Blue Lake Capital, BEYONDSOFT, Panggu Venture Capital, Zhaohui Capital, Caixin Capital, Daohe Long-term Investment, Yigao Capital, Mingde Capital, and others. The company attracted attention and participation from dozens of institutions, with the next financing round currently in final stages.

This signals both strong capital validation of Deep Insight's technology roadmap and implementation capabilities, and a broader industry trend: pragmatic implementation paths driven by foundational original technology are becoming the core development direction for physics AI.

In late 2024, founder Chen Kai first defined the "human learning" original technology roadmap. On October 10, 2025, the company released the "By Humans, Beyond Humans" General Embodied Intelligence Roadmap, establishing a three-pillar core strategy: "human data as the starting point, motion modeling as the center, robots designed for AI."

From perceiving and recording the world, to understanding physical laws, to executing real-world tasks—Deep Insight's full-stack self-developed system establishes a complete zero-to-one viable pathway for physics AI scaling.

Data is physics AI's core iteration fuel and a long-standing industry bottleneck. Simulated data is cheap but diverges from reality; real-robot data is high-quality but expensive and limited in scenarios—neither can sustain general foundation model training.

In late 2024, Chen Kai established the "human data as the starting point" technical direction. In June 2025, the company finalized its first-generation self-developed head-mounted data collection device, featuring a modular ergonomic design combining head-mounted sensing and a waist-worn main control unit, launching large-scale real-world data collection. Its design approach and deployment pace precede U.S. Scale AI's comparable equipment by nearly a year; its first-person dataset planning preceded NVIDIA's publicly announced EgoScale project.

After refinement across hundreds of real-world scenarios, the device achieved significant improvements in wearability and collection efficiency. While the industry only recently focused on first-person data, Deep Insight has completed multiple hardware iterations and scenario validations.

The company has established tens of thousands of hours of DeepAct multimodal first-person human data, spanning more than a dozen Chinese cities and hundreds of real-world scenarios. Deep Insight views genuine general-purpose embodied AI foundation models as requiring tens of millions of hours of high-quality human data for training. The company is simultaneously advancing million-scale computing infrastructure and planning tens of millions of hours of high-quality human data reserves.

Pursuing this goal, Deep Insight recently established a strategic partnership with Insta360, leveraging its expertise in panoramic imaging and spatial perception to accelerate physics AI's foundational data ecosystem scaling.

Centered on "motion modeling as the center," Deep Insight built a multidimensional coordinated model technology matrix spanning physical common sense, spatial reasoning, world models, and dexterous manipulation, comprehensively leading across seven major international benchmarks.

In March 2026, Deep Insight released PhysBrain 1.0 Embodied General Intelligence Foundation System, released one week before U.S. Generalist AI's GEN-1, with its core technology framework publicly cited by Physical Intelligence's flagship model π0.7.

The team recently achieved a major zero-shot generalization breakthrough: without specialized real-robot training or simulated data annotation, robots can autonomously complete novel tasks using only pre-training capabilities from human operation data. This marks the first domestic demonstration of the complete human-data-to-real-robot-execution pipeline under a fully self-developed technology system, proving human data can both help robots "understand the world" and directly translate into execution capabilities, clearing core obstacles for physics AI cost reduction and deployment.

Following "robots designed for AI," Deep Insight designs every aspect of hardware—structural proportions, degree-of-freedom configuration, control logic—specifically for AI models, achieving high-precision, low-latency motion mapping through deep software-hardware integration.

In September 2025, the company released Prime, a full-size humanoid robot and the world's first industrial-grade fully autonomous humanoid capable of independent standing. In February 2026, it globally premiered high-degree-of-freedom humanoid pipette manipulation demonstrations, showcasing continuous human-like manipulation capabilities, attracting widespread domestic and international attention.

Leveraging the human learning roadmap, Prime innovatively achieves direct motion output from pure human data without real-robot fine-tuning, establishing the complete "data training-model inference-real-robot deployment" pipeline and globally pioneering the "Robot for AI" technical closed-loop.

Building on this foundation, Deep Insight expanded its product matrix: Prime U, a wheeled dual-arm humanoid achieving millimeter-level precision and millisecond-level latency; Prime Lite, a lightweight 3D-printed robot for education and developer use, substantially lowering adoption barriers.

Deep Insight pursues a "restrained yet flexible" pragmatic commercialization approach: core resources concentrate on long-term foundation models, while intermediate R&D results simultaneously undergo market validation, creating positive feedback between technical development and scenario deployment.

All product lines—data collection devices, educational robots, teleoperation systems—have achieved commercial contracts, accumulating tens of millions of yuan in orders, making it one of physics AI's fastest full-product-line implementers. Products emerge from genuine R&D needs; customers proactively engage upon release with strong market demand.

· Scientific research: Full-size humanoid robots entered top university biology labs, enabling high-precision flexible experimental operations and exploring AI-assisted scientific discovery. Self-developed data collection devices partner with multiple research institutions and industrial scenarios for workflow documentation, operation guidance, and skill preservation.

· Education and developers: Prime Lite provides complete instruction spanning hardware design, 3D modeling, and model deployment as a physics AI platform for the intelligent era. Currently partnered with top domestic secondary schools and universities for real-world demonstration courses, with coverage by China Education News and Science and Technology Daily, and market expansion into Australia and other overseas markets.

Deep Insight is the first embodied general intelligence company jointly incubated by Zhongguancun Academy and the Zhongguancun Artificial Intelligence Research Institute, with its development consistently aligned with national strategic science and technology priorities. In June 2026, it appeared successively on CCTV's "Morning News" and "Evening News," exemplifying collaborative innovation between academia, industry, and research.

The company is led by founder Chen Kai and co-founders Zhang Yibo and He Xuguo. Chen Kai and Zhang Yibo, both from USTC's Gifted Youth Program and former roommates, bring long-term AI and large model expertise and physics/scientific intelligence background respectively; years of collaboration formed their stable management core. He Xuguo, with over a decade in humanoid embodiment, strengthened the company's capabilities in humanoid design, engineering, and rapid iteration. As technology and business advance, the team attracts global top talent; core members now span large models, multimodal systems, world models, robot hardware, and systems engineering—a "hexagon" full-stack team rare in physics AI.

Founder Chen Kai stated that physics AI development will successively progress through three stages: "human data large-scale pre-training, embodiment data feedback iteration, real-world autonomous evolution," with the industry currently in the critical first-stage window.

This financing round will deepen physics AI full-stack technology and the human learning roadmap. The company will anchor on "building globally leading physics AI foundation models," steadily advancing technological iteration and scenario implementation along the difficult but correct long-term path, driving domestically developed physics AI toward genuine industry adoption.