Robot planning often mixes continuous motion with discrete decisions such as contact modes or primitive actions. Most existing systems rely on symbol sets that are hand-designed or derived from language supervision, without asking whether those symbols actually make the underlying optimization easier. In this single-source study, Pan, Liu, and Mao take an inference-algorithm-centric view: a good symbol is one that exposes the discrete commitments that simplify solving the original planning problem.
The authors introduce the concept of solver-sufficient symbolization and show that, for finite-horizon piecewise-affine systems, this property can be characterized exactly. They then develop a counterexample-guided algorithm that invents such symbols automatically, and extend the principle to optimization objectives parameterized by general neural networks as energy functions. This allows solver-sufficient symbols to be induced from continuous optimization landscapes rather than from human intuition.
In experiments on simulated and real-robot tasks, the invented symbols reduced solving cost while preserving solution quality, and also improved interpretability and steerability. The theoretical results also offer an explanation for why certain human-defined symbols are computationally useful. The paper is a single contribution, so there are no conflicting sources to compare; its significance lies in shifting the question from what symbols mean to what symbols do for the solver.