At the 2026 Yunqi Conference, three embodied AI founders offered the same assessment: technical paths are still diverging at an accelerating pace, and the industry’s winners will be determined by who can cycle data, models and real-world feedback faster. What deserves attention is that embodied AI is moving beyond a stage centered solely on comparing model capabilities.
What this is
Embodied AI is a technology system that enables AI to perceive its environment and perform actions through robots and other devices. This discussion covered three parallel paths: VLA (connecting vision, language and action), world models (predicting changes in the environment), and teleoperation-based data collection. FIMOI Technology, Noematrix and Manifold Space have made different technical choices, but all three believe that at this stage, no single solution can solve everything.
The common direction proposed by all three companies is to establish a “data–model–deployment–feedback” loop: robots operate in real-world settings, generate new data, and then use that data for training and iteration. Compared with laboratory demonstrations, such systems depend more heavily on training workflows, evaluation systems and engineering organization.
Industry view
The industry is forming a limited consensus: real applications expose problems more effectively than evaluations in closed environments. Physical-world data collected through teleoperation offers a high degree of certainty, while synthetic data generated through simulation and world models is still constrained by realism, success rates and quality-control costs. This does not mean simulation has lost its value. All three guests believe it is better suited to augmenting datasets or assisting reasoning than to immediately replace real-world data collection.
The risks are just as clear. The industry remains in a period of rapid route changes, and training methods, compute efficiency or data-mixture strategies that are underestimated today may soon become the primary bottleneck. Deployment and post-training costs may also keep moving closer to the front of the value chain. More critically, many companies currently have models and data but lack the engineering loop needed to operate them reliably. If real-world scenarios are insufficient or feedback quality is weak, even the most advanced models may remain stuck at the demonstration stage.
Impact on regular people
For enterprise IT: procurement for embodied AI projects may shift from a single model or piece of hardware toward data pipelines, evaluation, deployment and continuous iteration capabilities. Traditional IT architectures will need to align more closely with on-site operations.
For individual careers: in the near term, new jobs are more likely to emerge in data collection, robot training, on-site implementation and engineering evaluation than in general-purpose model R&D. The value of cross-disciplinary collaborators will also rise.
For the consumer market: related products will remain focused on pilots and specialized use cases over the next year. Mature services still have a distance to go before reaching mass adoption, and the pace of price declines should not be overstated.