Long-Horizon Planning
Overview
Long-horizon planning asks how an agent can make progress toward a distant goal when many actions, dependencies, and changing conditions separate it from success. Common approaches break tasks into persistent subgoals, use world models to predict future states, and combine learned policies with search, diffusion-based trajectory generation, or symbolic constraints. Language and vision agents also rely on memory, progress tracking, and replanning to connect instructions to executable actions in settings such as robotics, autonomous driving, household tasks, and tool use. Methods are improving, but prediction errors and poorly grounded subgoals can compound over time, while benchmarks still reveal a gap between performing individual steps and reliably completing an entire task.
Latest papers
60 papers10 papers in the last four weeks, up 67% on the four weeks before. 0.1% of all new papers.