Princeton's new algorithm solves assistance games in one round

In human-robot collaboration, robots often have to infer what you really want—a problem formalized as assistance games. These games are notoriously hard because optimal strategies require solving a POMDP (partially observable Markov decision process) online, which is computationally intractable for real-world tasks. Researchers Elle Lazarski and Jaime Fernández Fisac (Princeton University) identify a class of assistance games where "pragmatic-pedagogic" reasoning can resolve goal uncertainty in just one interaction step. This collapses the full-horizon planning problem into a tractable best-response computation, making the game exactly solvable without expensive belief-space search.

The key insight is that humans and robots can coordinate more effectively when the robot treats actions as communicative signals—not just task execution. The authors show that mainstream inverse optimal control (IOC) methods, which infer goals purely from observed actions, hit a fundamental "inference ceiling" because many distinct goals look identical under task-optimal behavior. Pragmatic-pedagogic reasoning, by contrast, lets the robot deliberately act in ways that disambiguate between competing goals, even if those actions seem suboptimal for the task itself. This mirrors how humans teach by showing, not just doing. The paper validates the approach on a simple collaborative block-building scenario, demonstrating that the robot correctly identifies the human's goal in a single round and assists effectively. The work will be presented at WAFR 2026 (Workshop on the Algorithmic Foundations of Robotics).