I develop embodied agents that reason about how their present actions reshape the future. Every interaction changes object arrangement, free space, and accessibility, and those changes accumulate: a robot can complete each task it is given while gradually leaving behind a workspace that makes every later task harder. My work formalizes the missing capability as anticipatory planning and builds it into planners we already trust. I wrote the long version up as a blog post: Robots that think about the future.
Most deployed planners are myopic: they solve each task in isolation and, lacking any model of what will be asked next, introduce side effects on future tasks which are not yet assigned. We introduce anticipatory planning objective that learns the impact of an action on future tasks while solving a current one. We validate our objective in a symbolic-level task planning domain.
Continuous parameters, such as an object's pose, determine whether an environment stays accessible or becomes cluttered and obstructed for future tasks. Anticipatory task and motion planning chooses these parameters to preserve access and reduce clutter over long-lived operation.
Much of the computation in task and motion planning is spent on samples that go nowhere. A placement can be feasible for the current action yet quietly shrink the options for every later action in the plan, triggering a cascade of failures and replanning. Using anticipatory planning cost as a prior over which samples to try first, the planner favors choices that leave room for the rest of the plan, reducing planning time and search effort without changing what the planner optimizes.