We focus on the design and development of algorithms to enable robust decision theory becomes practical computational tools for robotics and related domain.
Such computational tools will enable robots to design their own strategies, such as deciding what data to use, how to gather the data, and how to adaptively improve its strategies, so as to accomplish various tasks well, despite various modelling errors and types of uncertainty, and despite limited to no information about the system and its operating environment.

Active Projects (under construction)

Decision-making in Partially Observed Scenarios

Due to uncertainty in actions and errors in sensors and sensing, the state of an agent/system is only partially observed and never known exactly. Despite such uncertainty, the agent must decide what actions to do now, so as to achieve good long term returns reliably. The Partially Observable Markov Decision Process (POMDP) is a general and mathematically principled framework for solving such decision-making problems. Specifically, it quantifies uncertainty using probability distribution functions, and computes the best action to perform with respect to distribution estimate of the states, rather than single states. However, exactly due to its careful quantification of uncertainty, solving a POMDP problem is computationally hard. We have been developing methods and software to enable POMDP becomes practical. Below are some of our recent work in this domain.

POMDP Planning for Problems with Complex Dynamics

POMDP Planning for Problems with Large Discrete Action Space

On-line POMDP Planning Toolkit

Robust Robot Manipulation

Partially Observed Inventory Control

Decision-making in Adversarial and Partially Observed Domain

Incorporating Defender's Behaviour in Autonomous Pen-Testing

Learning and Planning

Better Generalization via Improved Forward Propagation Primitive

Past Projects

Trademarks for Object Fetching with no Real Data