Om AI's Zhao Tiancheng: Betting on Physics AI's 'Streaming' Future
Om AI, led by CEO and Chief Scientist Dr. Zhao Tiancheng, is pioneering 'edge-native streaming multimodal' AI with its VLX model series. This architecture, designed from the outset for edge computing constraints, enables a complete closed-loop system for physical AI, integrating continuous perception, precise localization, and action decision-making. Unlike the prevailing trend of 'offline frame extraction' in digital AI, Om AI's approach treats video as a continuous stream, allowing AI to constantly process information without explicit prompts. This 'streaming' paradigm, which Zhao Tiancheng has championed since 2019, contrasts with the industry's more fragmented approaches to physical AI, such as VLA, video generation, 3D simulation, and JEPA. The VLX series aims to equip physical terminals like robots, drones, and AI PCs with the cognitive capabilities to transition from passive command execution to active scenario adaptation. Om AI has achieved a commercial closed-loop from simulation to industrial application, boasting a model loop, data loop, and business loop with revenues in the hundreds of millions. The company emphasizes the necessity of multimodal training for generalization in the open physical world, a principle validated by their early success with surveillance data, which outperformed models trained on years of specialized data. Om AI's strategy focuses on mature hardware first, such as PCs and IoT cameras, before expanding to emerging terminals like drones and robots, leveraging their existing expertise for new applications.
Om AI's focus on 'streaming multimodal' AI for physical applications represents a contrarian bet against the dominant 'offline' processing paradigms. By prioritizing continuous data flow and edge-native design, the company aims to achieve superior real-time perception and decision-making crucial for physical interactions. This approach addresses a potential bottleneck in current AI development, where the transition from digital to physical domains often encounters latency and generalization challenges. The emphasis on a complete closed-loop system—encompassing perception, localization, and action—highlights a systems-level perspective, acknowledging that true physical intelligence requires more than isolated model capabilities. Om AI's strategy of targeting mature hardware first before moving to emerging platforms suggests a pragmatic approach to market penetration, seeking to build a data flywheel and establish a competitive moat through early adoption and iterative learning. The long-term viability of this approach will depend on its ability to scale efficiently and demonstrate tangible advantages over more conventional, cloud-centric AI solutions in diverse physical environments.
AI-generated to prompt reflection — not editorial opinion, not advice, not a statement of fact. How this works.