Why robots are stuck in demos? iFlytek's new company gives answer: lack of 'body cognition'

iFlytek has quietly registered a new company, Anhui Yaofang Intelligent Technology Co., Ltd., with a registered capital of 300 million yuan, led by Pan Jia, a chief scientist of iFlytek and key contributor to the National Science and Technology Progress First Prize. The company focuses on building the 'brain' for robots, arguing that the current mainstream VLA (Vision-Language-Action) architecture is not the final solution due to its lack of consequence prediction. They propose that world models need an additional component: 'body cognition' — the ability for a robot to understand its own physical limits. The team released a technical report, iFLYTEK-Embodied-Omni, which introduces a unified multimodal embodied foundation model that combines VLM (Vision-Language Model), VGM (Video Generation Model), and AGM (Action Generation Model) with a shared multimodal self-attention mechanism. The model includes inverse kinematics as a training task to enhance body cognition. In benchmarks, it achieves an average success rate of 89.6% in LIBERO-Plus zero-shot generalization and over 93% in RoboTwin 2.0 dual-arm tasks. The company plans to integrate with iFlytek's subsidiary Lingdong General Robot to deploy the brain in real robots.