Quoting Francesco Farina:

The firehose of high-dimensional, real-time sensor data (vision, lidar, tactile, auditory, proprioceptive) from a world populated by embodied agents will dwarf the entire current internet dataset. Attempting to funnel this raw sensory stream back to central servers for storage and processing is not just economically daunting, it’s likely physically impossible at scale. The world simply generates too much information, too quickly.

The time it takes for sensor data to travel to a central brain, be processed, and return as an action command is often an eternity when dealing with physical dynamics. A drone adjusting to a sudden gust, a robotic surgeon countering a tremor, a logistics bot avoiding a collision – these demand intelligence that lives at the point of action. Centralized processing, even with 5G or 6G, faces irreducible latency limits imposed by physics itself.

Centralized systems are inherently brittle. Server outages, network disruptions, cyberattacks, even software bugs in the core model can cascade, paralyzing entire fleets, factories, or cities dependent on that single brain. Robustness in the face of failure, damage, or unforeseen circumstances demands redundancy and graceful degradation – hallmarks of decentralized systems. Nature doesn’t run on a single server farm.

True mastery of the physical world requires continuous adaptation to local, specific conditions. A robot arm’s optimal movement strategy subtly changes as its joints wear. Grasping success depends on the unique friction coefficient of this object, right now. Learning purely from fleet-wide averages pushed down from a central model is too slow, too coarse. Agents need the capacity to learn in situ, refining their models based on their own unique experiences and immediate environment. Centralized training struggles to capture this granular, continuous, embodied learning.