My PhD work: robots that think about the future
Imagine you bring a robot home
It's a capable machine. It can see, drive around, pick things up and set them down again. It knows what a bowl is. You give it one job, in plain language:
“Clear the dining table.”
On the table right now: a clean mug, two clean bowls, and one dirty cup from this morning.
It does exactly what you asked
When it finishes, the table is bare. Job done — and here is where everything ended up.
The table is clear. The robot was asked to clear the table, and the table is clear. By every measure it was given, this is a complete success.
It is also the reason you can't work at your desk or cook dinner.
Nothing went wrong, and that's the problem
The robot didn't malfunction. It didn't misunderstand you, drop anything, or take a strange route. It was told to clear the table, it found the cheapest way to clear the table, and the cheapest way happened to be “put each thing on the closest surface that has room.”
The trouble is that the cost of that choice doesn't land on the robot. It lands on you, later, and on every job that comes after. Want coffee? Move a bowl first. Want to cook? Clear the stove first. Want to wash the dirty cup? It's across the room on a burner, and the bowl beside it has to move before you can reach it.
None of that appears anywhere on the robot's report card, because the report card closes the moment the task ends. It gets full marks, hands the mess to the future, and waits for the next instruction.
Do this for a week and you get a home that quietly becomes harder to live in — every individual decision defensible, the accumulated result a mess. Humans avoid this without thinking. You don't put a dirty cup on the stove, not because someone forbade it, but because you have a rough sense of what the kitchen is for and what you'll want from it next.
Teaching robots that sense is the problem I work on. It has a name: anticipatory planning.
Now let it weigh the future
So change what the robot is trying to be good at. Instead of scoring itself only on the job in front of it, let it also account for what the next jobs will cost, given the state it is about to leave behind. That is the only change. No rules, no list of which object belongs on which shelf.
Watch where the dishes end up now.
The desk is yours again and the stove is usable, and notice that nobody arranged that. The robot was never told that a cabinet is for dishes or that a stove is for cooking. It worked out that parking a bowl on the stove is a decision it would have to undo, and that the cabinet is somewhere things can sit without getting in the way of whatever comes next. The habit looks like tidiness. It is really just foresight.
But look inside the cabinet. It got the big call right — that cabinet rather than the nearest counter — and then dropped things wherever there was room, in the order it happened to carry them. The foresight ran out at the cabinet door.
This is the part that surprised me when I started working on it. Choosing which place is not enough. Where exactly, within that place is a second decision, and it is the one that decides whether the cabinet stays a cabinet or turns into a jumble.
Both decisions, at the same time
So the robot has to make two kinds of choice together: the categorical one — which surface, which container, which order — and the physical one — which precise spot, which pose, how much room to leave beside it. Decide them separately and you get the cabinet above: a sensible plan executed into an awkward arrangement.
Look at where the dirty cup went. Nobody said “wash the cup.” The instruction was only to clear the table. But a dirty cup is going to need washing sooner or later, and the sink is where washing happens, so leaving it there costs nothing now and saves a trip later. That behavior wasn't programmed in. It's what falls out of a planner that is weighing the future at all.
The idea in one line
Here is the entire technical content of my thesis, in words. A normal planner tries to make one number as small as possible: the cost of doing the job in front of it. An anticipatory planner minimizes a sum of two things instead.
What an anticipatory planner makes as small as possible
That's it. Now + future. The first term is what every planner already optimizes. The second is the one that's missing, and adding it is what turns “put it on the nearest surface” into “put it where it won't be in the way.”
The word expected is carrying weight there. The robot doesn't know what you'll ask next. It's working with a sense of what tends to get asked — bowls get used for cereal, dirty cups get washed, desks get worked at — and weighing each possible future by how likely it is. You're not predicting tomorrow. You're avoiding choices that are bad across most plausible tomorrows.
And the two terms genuinely pull against each other. Walking to the sink costs more than setting the cup down right here. Anticipatory planning spends a little more now to save more later, which means its first move can look like the slower one. Here's that trade with real numbers from one of our experiments:
That was two tasks. The more useful question is what happens over a long run — ten tasks in a row, averaged over 64 sequences in the cabinet kitchen.
Both planners start in the same tidy kitchen, so the first task costs them the same. After that they come apart, and the shape of the two lines is the whole argument. The myopic planner bounces around the same level for all ten tasks: every job costs about what the last one did, because it keeps handing itself a kitchen in roughly the state it found it. The anticipatory line heads downward. It is not getting better at picking things up — it is getting better at the room. Each task leaves the kitchen a little easier to work in, and the next task collects the benefit, so the gap between the two lines is wider at the end of the sequence than at the start.
That gap is what the second term buys you, and it only shows up if you measure a whole deployment. Watch any single task and the two planners look comparable — occasionally the myopic one looks better.
How can it possibly know the future?
It can't, and it doesn't try to. What it has is a learned instinct for which arrangements tend to work out well.
We build that instinct the slow, honest way. Put the robot in a simulated kitchen, hand it a long sequence of ordinary tasks, and record what everything actually cost. Do it hundreds of times. A pattern emerges: states where things are grouped by kind, where surfaces are left open, where the thing you'll need next is reachable, are states where the next task comes cheap. Nobody labels those states as “good.” The costs do the labeling.
The piece that learns this is a graph neural network, and the reason is almost aesthetic: a kitchen already is a graph. Each object, surface and container is a point; the relationships between them — this bowl is on that table, this mug is in front of that bowl, this cabinet is next to the sink — are the lines connecting those points. A graph network is a model built to read exactly that shape of data, so instead of flattening the room into a list of coordinates and hoping the model works out what matters, you hand it the room as a web of relationships and let it learn which patterns in that web mean trouble later.
It's worth being clear that none of this replaces the planner. The learned part offers an opinion about the future; the planner still does the planning and still guarantees the plan is valid and executable. That split is deliberate. I want the common sense to be learned and the correctness to be proven, rather than hoping one model does both and is honest about it.
Why I think this matters
Robots are moving out of the single-demo world — one tidy task, one reset, one score — and into homes, warehouses and hospitals where nothing resets. In that world, the state a robot leaves behind is the state it has to work in tomorrow, and so does everyone else sharing the space.
Which makes “did it finish the task?” the wrong question, or at least an incomplete one. The question I'd rather we asked is whether the place is still good to work in afterwards. A robot that finishes every job and slowly makes your kitchen unusable has not, in any sense that counts, succeeded.
The benchmarks mostly don't measure this yet. They reset the room between tasks, so the long-term consequences are invisible by construction, and a short-sighted robot can score perfectly. Changing what we measure is, I suspect, the part that will actually move the field.
The work behind this: ICRA 2023 introduced the anticipatory planning objective, RA-L 2026 carried it into task and motion planning for cluttered, persistent spaces, and an IROS 2026 workshop paper found that the same learned estimate also makes planning itself faster. The formal version of all of it is in my research statement, and the simulations here were built in PyBullet with a Fetch robot.