Next wave of AI will be physical

The year 2025 marks a definitive inflection point for Artificial Intelligence, transitioning from digital-only software to "Physical AI" and "Embodied AI" systems capable of perceiving, reasoning, and acting within the real world.

This evolution is driven by the development of "World Models" and embodied foundation models, which function as internal simulators to predict environmental dynamics and plan complex actions without relying solely on costly real-world interaction.

To address the persistent "Sim-to-Real" gap, researchers are deploying hybrid strategies that combine high-fidelity online imitation pretraining in simulation with offline finetuning on sparse real-world data.

Major infrastructure platforms, such as NVIDIA Cosmos, are accelerating this progress by providing generative world foundation models to synthesize the vast amounts of physical interaction data required for training.

Consequently, the strategic focus of the industry has shifted from merely scaling computational parameters to optimizing software architectures that can integrate multimodal perception with resilient control policies.

Commercial adoption is surging in industrial sectors, where autonomous mobile robots and humanoids are replacing rigid legacy automation, though widespread deployment is still constrained by hardware limits and safety requirements.

To overcome these barriers, future systems are prioritizing adaptive resilience and "learning from experience" over absolute precision, allowing robots to recover from errors in unstructured human environments.