Should robots separate what they know about the world from what the task asks?

Robot policies often fail when environments, teammates, or constraints change, because they entangle world knowledge with task logic. This work argues that separating these—treating world properties (physics, embodiment) as fixed and task logic (goals, constraints) as modular—is the right structural choice. The team formalized this through Bayesian evidence and built AICON, a differentiable graph of estimators that learns world dynamics without task-specific data, paired with a learned policy that modulates gradient flow. Tested on three heterogeneous robot problems, the approach outperformed end-to-end baselines and zero-shot transferred to real hardware.