A computational framework for understanding embodied plans

This research presents GenEA, a novel agent representation that reasons over multi-step, long-horizon embodied plans by coupling coarse biomechanics of an articulated body with interactions using simulation-based intuitive physics. Embedded in a generative model, GenEA yields a joint probability distribution over goals and embodied plans. Experiments across object manipulation and navigation show that GenEA recapitulates human reasoning about embodied plans through simple transformations of this distribution and consistently outperforms strong alternatives including plausible heuristics, standard task-and-motion planning, and state-of-the-art vision-language models.