A few findings stood out:

📍 Current research is dominated by assembly-targeted use cases — mostly on stationary cobot arms. More complex morphologies (humanoids, mobile manipulators) remain rare.

📍 Vision and language dominate the inputs, while tactile and force feedback — arguably essential for, e.g., real-world manufacturing use cases — are almost absent (just 4 of 42 studies).

📍 "Foundation model paradox": LLMs and VLMs are increasingly used as modular, swappable planners. This is practical, but it pulls systems away from the tightly coupled sensorimotor intelligence and closer to "Physical AI."

📍 True Vision-Language-Action (VLA) models stay rare, bottlenecked by the scarcity of domain-specific kinematic training data.

Takeaway: Industrial EAI is still early. Much of the work realizes the label but not the embodied principle — intelligence emerging from the coupling of body, learning, and environment. The near-term path forward is likely to be task-specific, human-centered systems, richer multimodal feedback, and real industrial benchmarks, not fully autonomous general-purpose robots just yet.

#EmbodiedAI #Robotics #Manufacturing #AI #FoundationModels #Humanoids