Coordination without command: active inference and the route to embodied intelligence
Frontiers in Neurorobotics · 2026-11-12 · 20
DOI: 10.3389/fnbot.2026.1898490 ↗Abstract
Embodied intelligence depends on bodies whose contact with the environment, distribution of load and capacity for action are only partly known and are changed by movement. We argue that this creates a practical reason to organize inference through distributed, hierarchical and plural processes, using active inference to relate local belief updating, contextual constraints and action selection within a common framework. The proposal brings together established ideas from factorized inference, distributed control and embodied robotics, distinguishing their mathematical description from the location of computation and authority over action. The octopus provides biological motivation through its combination of peripheral sensorimotor processing and central influences on behavior, while an articulated hexapod illustrates selected architectural choices through six leg-local predictive controllers coupled by neighbor messages and a shared mechanical body. The simulation includes intact locomotion and movement following programmed limb withdrawal and graph reconfiguration, although these observations do not isolate the contribution of inference or establish remapping. Together, the biological example and computational illustration clarify how local inferential processes may remain coordinated across bodily interfaces and timescales without a controller prescribing the complete joint trajectory.
EUVIMED