At the 2026 China International Fair for Trade in Services, a special session on Physical AI and Embodied Intelligence Robot Innovation Ecosystems was convened in Beijing on September 11. Kai Xuan Nie, founder and CEO of Songying Technology, delivered a keynote address, outlining the industry's trajectory. He asserted that as generative AI and intelligent agent technologies advance, artificial intelligence is accelerating its convergence with robotics and smart equipment, extending applications from digital spaces into the physical world.
Referencing industry leader Jensen Huang's forecast, Nie noted that Physical AI is poised to drive transformation across a manufacturing and logistics sector valued at approximately 50 trillion US dollars. This evolution opens new development horizons for software, hardware, chips, models, and sensors alike. However, he stressed that beneath this industrial opportunity lies a series of engineering challenges that must be addressed to materialize this potential.
Where to Begin: Conquering the Data and System Barriers
The journey of Physical AI into reality is gated by two primary hurdles: data sufficiency and system integration. On the data front, quantity alone is not the answer. Currently, the industry accumulates training data through robotic teleoperation, first-person perspective collection, and simulated generation. While each path offers distinct value, the existing datasets fall short in scale, quality, and richness when compared to the needs of models that must genuinely comprehend physical laws and possess real-time interaction capabilities.
To achieve this, Physical AI demands high-quality, multi-modal data that captures dynamic interaction processes. Every perception, judgment, and action of a robot is continuously linked to its surrounding environment, making the scalable generation of such interactive data a critical R&D challenge. Advances in computational power and algorithms now enable the generation of synthetic data at scale, which can expand dataset breadth and cover more scenarios, including edge cases. Real-world data, conversely, serves to calibrate and validate these synthetic sets, facilitating a feedback loop from digital environments back to physical devices. The synergy between synthetic and real data is essential for fostering a continuously evolving training ecosystem.
Beyond the data layer, robots entering a factory become part of a broader production system. They must integrate with existing industrial equipment, align with production rhythms, and comply with management system dispatch protocols. Industrial production demands strict accuracy, real-time responsiveness, and stability. A robot capable of performing a single action is far from ready to join a full production line; it must resolve connectivity issues across devices, software, data streams, and production processes. Therefore, the industrialization of Physical AI requires both enhancing individual machine intelligence and solving systemic coordination challenges.
A Unified Pipeline: ORCA OS Streamlines R&D and Deployment
To address these data and system coordination issues, Songying Technology has developed ORCA OS, a Physical AI operating system. Designed to cover the entire R&D lifecycle, ORCA OS connects scenarios, data, algorithms, training, evaluation, and deployment within a unified environment. It offers high-fidelity physical simulation and data synthesis capabilities while also supporting model training, validation, and integration with real industrial settings, thereby preventing R&D phases from operating in silos.
A key subsystem, ORCA Connect, is dedicated to bridging existing industrial software and production data, with compatibility for mainstream platforms like Dassault and Siemens. This allows robot training, scheduling, and deployment to integrate smoothly into established industrial workflows. In terms of scenario construction, the system leverages AI to generate and assemble 3D environments or replicate real-world settings through scanning and modeling. These scenarios serve not only as visual representations but also act as training grounds for robot perception, action, and evaluation. The simulation engine supports multiple physics fields, including rigid bodies, soft bodies, and fluids, allowing different materials and phenomena to interact within the same digital environment. In April 2026, the company released a real-time multi-physics coupling simulation demo, showcasing simultaneous interaction between rigid, soft, and fluid bodies.
Furthermore, ORCA OS simulates various sensors—such as vision, radar, inertial measurement units, force, and tactile sensors—to generate multi-modal data essential for robot understanding. The system currently produces data across more than 20 dimensions. While current models may not fully utilize all these dimensions, this rich data reserves capacity for future training and research. Regarding training efficiency, the value of simulation lies in parallel computing, allowing models to undergo more extensive training and testing before encountering physical hardware. Real-world situations that are difficult to reproduce frequently can be iterated upon repeatedly in digital environments, and high-cost tests can be pre-validated in simulation. Evaluation systems subsequently break tasks into segments, providing feedback on model and algorithm performance. This creates a closed loop where scenario construction, data generation, model training, and results evaluation are seamlessly connected rather than disparate steps. Since its inception in 2021, Songying Technology has pursued this objective, completing seven successful funding rounds.
An Open Architecture: Connecting Compute Resources and Developers
The advancement of Physical AI requires contributions from diverse technologies, including compute power, industrial software, robotic bodies, model algorithms, and real-world scenarios. Songying Technology has chosen an open and decoupled technical architecture to foster this ecosystem. This approach connects various models, robots, and applications at the top layer, and supports a broad range of computing power at the bottom, empowering enterprises to select their preferred technology stack. Beyond mainstream international GPUs from Nvidia, AMD, and Intel, ORCA OS is compatible with domestic chips such as MThreads, MetaX, Biren, Iluvatar CoreX, and Lilith. This openness not only offers more choices in compute but also integrates disparate technologies, tools, and devices into a unified Physical AI R&D pipeline.
Built upon this foundation, the company offers ORCA Lab, a free developer edition designed for educators, students, individual developers, and innovative teams. ORCA Lab provides tools for scenario construction, data synthesis, training, and evaluation. To encourage broad participation, its hardware requirements are lowered to an RTX 3060 level, enabling developers to initiate Physical AI tasks from their personal laptops. Currently, ORCA Lab boasts a professional developer base numbering in the thousands, creating more possibilities for connecting scenarios, data, models, and applications within the same tool environment.
From Demonstrations to Industrial Reality
Ultimately, Physical AI must solve real-world industrial problems, and its capabilities are only validated when deployed in specific contexts. At the World Artificial Intelligence Conference, Songying Technology expanded its demonstration of multi-agent coordination systems. Humanoid robots, AGVs, tracked vehicles, and drones executed collaborative tasks within a single simulation environment, with provisions for human intervention to test model resilience against random changes. Following the ORCA OS release, the innovation received in-depth coverage from several major media outlets including People's Daily, Xinhua News Agency, China Daily, Science and Technology Daily, and CCTV.
Focusing on actual needs in manufacturing, electric power, and other fields, the company is fostering scenario-based competitions and industrial collaborations. These initiatives bring together robot enterprises, scenario owners, developers, and technology platforms for joint training and validation. Such mechanisms provide robot companies with access to real tasks and training data while helping scenario owners select appropriate technologies and partners. In May 2026, a robot scenario application challenge was held in Hangzhou, and an embodied intelligence competition focused on power scenarios is underway with 150 participating units. An international embodied intelligence event is planned for Shanghai in October 2026. These competitions connect real tasks, scenario data, training validation, and partner selection, subjecting technology to rigorous testing against specific industrial requirements.
In conclusion, for Physical AI to transition from singular capabilities to systemic synergy, stronger models are essential, but so is coordinated interaction among scenarios, data, compute, training, evaluation, industrial software, and physical hardware. Only through the mutual verification of synthetic and real data, and the seamless connection between digital training and real-world deployment, can models truly move from digital environments to physical devices. Through ORCA OS, Songying Technology aims to facilitate this transition, accelerating the arrival of Physical AI in the real world.