iFlytek Spins Off 'Yaofang Intelligent' to Build Robot Brains, Argues VLA Is Not the Final Answer

On the eve of WAIC 2026, iFlytek quietly registered a new subsidiary — Anhui Yaofang Intelligent Technology Co., Ltd. — with 300 million yuan in registered capital, led by Pan Jia, a principal winner of the National Science and Technology Progress First Prize and iFlytek's distinguished scientist specializing in embodied AI.

The company's core thesis is contrarian: the current mainstream approach to building robot brains is fundamentally flawed, and VLA (Vision-Language-Action) is not the endgame. Pan Jia argued that robots on exhibition floors move slowly not because of hardware limits but because "they're afraid of making mistakes." Speed can be increased, but error rates rise and demonstrations fail. The real bottleneck is that robot brains haven't crossed the threshold of practical usability.

Yaofang contends that VLA models are merely reactive systems that handle the present moment but cannot anticipate consequences. The emerging world model (WAM) approach improves on this, but the company identifies two hidden problems: error accumulation from serial prediction-to-action pipelines, and a lack of "body cognition" — the model can predict the world but doesn't understand its own physical limits.

To address these gaps, Yaofang released the iFLYTEK-Embodied-Omni technical report, proposing a "dual-core brain" architecture. It adds a Video Generation Model (VGM) alongside the traditional VLM to predict future visual states before action execution, and incorporates inverse kinematics as a dedicated training task for self-body awareness. On the LIBERO-Plus zero-shot benchmark, the system achieved an average success rate of 89.6%.

Yaofang's data strategy also diverges from the industry norm: nearly half of its training data consists of spatial reasoning and task planning data rather than expensive real-robot data. The company argues that deploying robots at scale before crossing the usability threshold would fail because users won't adopt unreliable products, breaking the data flywheel before it can spin.

The new venture builds on iFlytek's decades of AI engineering experience, evolving from speech recognition to the "Super Brain 2030" plan and now to embodied intelligence. Yaofang provides the AI foundation while iFlytek's subsidiary Lingdong General Robotics handles the full industrial embodied intelligence pipeline.