Computational models to predict user trajectories in dynamic environments / Modèles computationnels pour la prédiction de trajectoires d'utilisateurs dans des environnements dynamiques

target selection, VR, optimal motor control, motor learning, HCI

Context of the project: Reaching an object (e.g selecting a 3D object in VR or an icon on the
desktop) is one of the most fundamental tasks in Human Computer Interaction (HCI). In HCI, Fitts’
law has been extensively used to predict the pointing time depending on the distance and size of the
target (object). It has been used to compare different devices, as well to develop advanced interaction
techniques. However, Fitts’ law remains a behavioural model providing little explanation regard-
ing the cognitive processes and thus it does not explain/predict how users adapt their behaviour
in dynamic environments e.g., tasks involving external forces or dynamic mappings between physical
and virtual movements. A model that would predict human produced trajectories in dynamic environments would inform the design of many non-static input-output mappings (e.g., adaptive mice, VR techniques that manipulate the mapping), by allowing counterfactual reasoning.
Objectives The objective of this internship is to understand how people produce and adapt their
trajectories in a new and/or dynamic environment. We embrace a model-based view of action, where
human policy builds on predictions of an internal world model of the task to be accomplished, in line
with the optimal control framework pioneered by Todorov. In this classical framework, the
internal model is static and identified beforehand. We hypothesise that, rather than being static, this
internal model is continually kept up to date, in light of conflicting prediction and sensory information.
Modelling and integrating this learning process in the optimal control framework is the open problem
that we address. To achieve this, we will adapt Todorov’s classical model, by having the internal model
inferred. This inference will be achieved by progressively updating the original outdated internal
model, based on an error signal between predicted and observed outcome. The rates of updating (how
often the model parameters are updated and by how much) will be determined from empirical data
that we already have.

Required Competencies
* Python
Optional Competences
* State space systems, optimal control
* experience with VR systems

Julien Gori
Gilles Bailly
Référent Universitaire: 
Fichier Descriptif: 
2 023

User login