Four Vectors
We develop world models for robots that work alongside people: models that let a robot (1) track a dynamic, partially observable world, (2) infer what each person believes about it, (3) notice what each person has missed, and (4) help at the right moment.
1 Mental models
People act on what they believe, not on what is true. Usually the two agree. When they do not, a person reaches for a mug where it used to be, unaware that someone has moved it.
Mental model theory holds that decision making, situation awareness and behaviour planning all draw on an internal simulation of the world: a belief state. In a team, each person also keeps a model of what every teammate believes and can do. This team model lets people work together fluently without explaining every step.
A robot can model every object in a room and still not know which ones its teammate has seen. To work with people, a robot needs a team model too.
We develop a core part of that team model for robots: a person’s belief state, or level-one situation awareness [1], which is what they believe about each object, such as where it is. We study it in dynamic, partially observable environments, where objects move and no one sees everything.
Following [1], the robot keeps its own belief state, βrobot, from what it observes, and infers each person’s, βhuman, from what that person could have seen. Where the two differ, the robot knows what the person has missed, and when to help.
- where the robot observes it, βrobot
- where the person believes it is, βhuman
- the person and their view
- a belief that is out of date
2 Current work
Our current work has three parts.
Belief-state ground truth. As [1] notes, no dataset yet records people’s situation awareness in a representative household or workplace, and one is needed to align inferred belief states with what people actually believe. We are building it. Participants carry out everyday tasks while fixed cameras stand in for the robot’s view, eye-tracking glasses record where they look, and short spoken questions record what they believe.
Whole-body demonstrations. We record people doing mobile manipulation tasks, using a lightweight wearable and fixed cameras, as demonstrations for robot learning. We plan to compare this with capture from VR headsets.
Open research. We plan to release data samples and technical reports, so that others can check the work and build on it.
3 Publication
Our work builds on [1], a preprint now under review, written at the Georgia Institute of Technology. It infers a teammate’s belief state as a human-robot team moves through a household. It evaluates the method in a realistic household simulation, extends it to a real robot, and demonstrates an active-assistance task.
Jack Kolb, Aditya Garg, Nikolai Warner and Karen M. Feigh. Inferring World Belief States in Dynamic Real-World Environments. arXiv:2604.11020 [cs.RO], 2026.
4 How we work
We record only adults who have given written consent, in closed sessions with no bystanders. We blur faces before any release and store every recording encrypted.
5 Working together
A helpful robot models the world twice: as it is, and as you believe it to be.
We want to hear from researchers working on human-robot teaming, situation awareness or robot learning from demonstration, and from the research programmes that support that work. We welcome joint studies and comparisons of methods.
If this work is close to your own, write to us. We are also hiring for five positions.