Can simple reactive control handle conflicting robot goals?

Reactive control typically fails at multi-objective tasks because static encodings create local minima when goals conflict. This work exploits a graph-based world model to resolve conflicts on-the-fly: lower-priority objectives are projected into the nullspace of higher-priority ones, with priorities updated continuously from state. On planar pushing of non-convex objects, it reaches 100% success across 100 configurations—versus 0% for gradient descent and ~55% for diffusion policy—without needing demonstrations. The approach transfers directly to real robots.