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. Full details are in my research statement.
Most deployed planners are myopic: they solve each task in isolation and, lacking any model of what will be asked next, introduce side effects that impede future tasks. I augment the task objective with a learned estimate of the expected cost of future tasks, so a planner chooses among the many states that satisfy the current goal the one that leaves the environment easiest to work in.
Where an object goes matters as much as which object moves. Loading a cabinet so that mugs occlude the bowls satisfies the task but forces costly rework later. I bring anticipation to task and motion planning, estimating how continuous placements affect plausible future tasks so that reachability is preserved and obstructions are avoided across long deployments.
Real deployments are shared. A staging spot that is convenient for one agent can be another's blocked retrieval path, and a robot that is efficient on its own terms can be inconsiderate to the people around it. I extend anticipatory planning to persistent shared environments, where the cost a decision imposes on others is part of the objective.
Anticipation should not require throwing away decades of planning research. I integrate learned anticipatory cost models directly into the Fast Downward stack via C++, PyBind, and PyTorch, so neural estimates compose with standard symbolic heuristics and the planner keeps the guarantees it already provides. The same estimates also act as a learned prior that reduces search and sampling effort on a single problem.
Anticipatory planning needs a distribution over future tasks and a goal specification, and the real world hands us neither. Instructions like "tidy up the kitchen" leave goals, objects, and acceptable placements implicit. I want to use LLMs and VLMs to build the components anticipation depends on — semantic world models, task-distribution priors, goal specifications — while decision making stays with planners whose long-term consequences can be verified rather than assumed.
How a robot grasps, places, or stacks often determines the state it leaves behind more directly than which task it chose. A placement that sets a pan down with its handle in a walkway succeeds at "place the pan" while creating a hazard. I plan to train VLA policies against objectives that reflect downstream environmental quality, distilling anticipatory preferences discovered by high-level planners into low-level skills.
Persistent deployment is not only the setting that demands anticipation, it is also the data source for learning it. A robot in the same home for weeks records which configurations stayed favorable and which repeatedly caused obstruction. I want agents that close this loop: refining future-cost models and updating task-distribution beliefs online from their own experience.
Warehouses, hospitals, and homes contain multiple robots and humans whose actions jointly reshape the environment. I am interested in anticipatory planning for this setting, where the future task distribution an agent must anticipate includes the tasks of other agents, and in the benchmarks the field will need to measure long-lived performance honestly.