Compositional energy models: learn reusable evaluation factors, compose them for new tasks

du_yilun · x · 2026-09-09

Using the example of placing a mug on a crowded table, the author argues that instead of directly predicting the placement, one can learn reusable local evaluation factors — support, clearance, reachability — each judging a local relationship rather than the whole scene; inference composes these factors into an evaluator and searches for a placement satisfying all constraints together. In their compositional-energy work, factors learned on small SAT problems compose to solve larger ones. The principle extends to planning and reasoning: by combining familiar skills, reasoning operations and expertise, a system can construct both a solution and the reasoning process used to find it.

Related event: Du Yilun Proposes Reasoning-Driven Robot Generalization(3 posts)→

Original post →

More from Research

Research channel →