A Reflection on Embodied Intelligence
Over the past few years, embodied intelligence has attracted increasing attention, capital, and talent. More people now believe that robots may eventually enter ordinary homes and become our colleagues, partners, and companions. Yet we must also admit a less exciting fact: truly deployable solutions are still rare. A bubble may be forming, and a downturn may be approaching. Still, I believe the vision itself is too bright to ignore. The road is winding, but the future is bright.
Robotics is not a new playground created by the recent AI wave. It is an old and deep engineering discipline, with foundations no less rich than computer science or AI itself. However, both industry and academia today often pay too much attention to benchmark scores, flashy demos, fundraising narratives, and publicity. We seem to spend too little effort studying robots as robots, and too much effort importing paradigms from other fields.
This βimportismβ may be one of the most unhealthy trends in embodied intelligence today. If we simply apply discriminative or generative foundation models from other domains to robotics, without seriously considering the structure, constraints, and physical reality of robots, we may end up with fast-moving but shallow trend chasing. This happened in the VLA era, is happening in the WAM era, and may soon happen again in the agent era.
I also do not think blindly following these imported paradigms is necessarily a good strategy for ordinary academic teams. From the WAM era onward, the leadership of the paradigm has increasingly been taken by companies with stronger resources, infrastructure, and publicity channels. This makes traditional academic research even more important. Academia should not merely follow authority. It should challenge assumptions, explore new paradigms, and ask what is truly essential for robotics.
This is also why World Guidance means a lot to me. In many latent world model approaches, the latent space is still constructed around visual differences. But for robotics, this may not be fundamental enough. The latent space for action generation should not merely compress visual change. It should capture what truly matters for action.
WoG follows this belief. Instead of hand-designing a hidden space, or simply borrowing one from another paradigm, we let action supervision drive the formation of the condition space in a data-driven way. In other words, the objective of action generation itself tells the model what future information is useful. To me, this is a more fundamental and elegant way to think about world modeling for robotics.
The future of embodied intelligence will not come from chasing every new slogan. It will come from respecting robotics as a discipline, understanding its real constraints, and building original paradigms that can truly push the field forward.
The road is winding, but the future is bright.
#EmbodiedAI #Robotics #WorldModels #VLA #WAM
