Research Interests:

Back to home

New: On the number of local minima for SLAM as an optimization problem -- paper accepted by ICRA 2012,( paper, code)

New: Kalman Filter (KF) and extended Kalman Filter (EKF) without tears

Here I tried to explain KF and EKF without using too much probability theory (1D KF, KF, EKF).

New: 2D I-SLSJF MATLAB code (code and data sets, documentation)

· More 2D SLAM Data Sets

1. Simulation data set in the SLSJF TRO paper (data, local maps, SLSJF results)

2. Simulation 3343 loop data (data only)

· 3D SLAM Data Sets

1. Simulation 90 loop data (data)

2. Simulation 870 loop data (data)

Research interest 1: Simultaneous Localization and Mapping (SLAM)

SLAM is the process of building a map of an environment while concurrently generating an estimate for the pose of the robot. In traditional SLAM, Extended Kalman Filter (EKF) is used and the computational cost is quadratic with respect to the number of features.

New map joining algorithms for large-scale SLAM

· Paper (IEEE Trans. on Robotics, Vol. 24, No. 5, 1121-1130, October 2008, with Zhan Wang and Gamini Dissanayake)

Local submap joining is a strategy for efficiently building large-scale maps. By first building a sequence of local submaps and then combining them together, the frequency of a big map update step is significantly reduced. Recently, we have developed a number of new map joining algorithms. The MATLAB code is now made available to SLAM researchers.

Convergence properties and inconsistency issue of EKF SLAM (IEEE Trans. on Robotics, 2007, with Gamini Dissanayake)

In traditional SLAM, Extended Kalman Filter (EKF) is used to estimate a state vector containing both the robot pose (including location and orientation) and the feature locations.

Many people regard the SLAM problem for a small size environment using range and bearing sensors is a solved problem. However, the algorithm may provide inconsistent estimation due to linearizatoin errors (Tardos04,Frese05,Bailey05). The convergence properties are proved but the proof is only provided for the simple linear case (Dissa01).

In this project, we aim to prove the convergence properties of EKF SLAM for the general nonlinear case, and understand in depth how the inconsistent estimation comes from. The final goal is to provide a robust and consistent SLAM algorithm which can be applied in real time applications.

Decoupling localization and mapping in SLAM

(D-SLAM, Int. J. of Robotics Research, 2007, with Zhan Wang, Gamini Dissanayake)

In traditional SLAM, the state vector contains both the robot pose and the feature locations and the localization and mapping are performed simultaneously by Extended Kalman Filter (Extended Information Filter). Recently, we have shown that SLAM problem can be reformulated such that the mapping and localisation can be treated as two concurrent yet separated processes: D-SLAM (decoupled SLAM: a non-linear static estimation problem for mapping and a three-dimensional estimation problem for localisation).

Trajectory planning in SLAM

(Simultaneous Planning, Localization and Mapping (SPLAM), Robotics and Autonomous Systems, 2006, with Cindy Leung, N. M. Kwok, Gamini Dissanayake)

Many existing SLAM algorithms have been focused on the offline estimation of the robots and features locations. This paper suggests to use Nonlinear Model Precdictive Control idea to actively choose the observation points to maximize the information gain.

Research interest 2: Robotic Search and Multi Robot Cooperation

Multi-robots cooperative control (in target localization, IROS'05, with Cindy Leung et.al.)

For target localization problem, a centralized cooperative control strategy plus a very coarse exhaustive search among a few control options results in a near optimal trajectory

Optimal search for multiple targets (in a known build enviornment, EJOR, with Haye Lau et.al.)

In a search and rescue scenario, given the structure of a building, the probability of the presence of multiple targets, how to decide the optimal search strategy to minimize the expected time of finding the targets?

Research interest 3: Systems and Control

L-infinity bounded robustness for nonlinear systems (IEEE Trans. on Automatic Control, 2003, with Matthew R. James)

This work generalized the Hill-Moylan-Willems framework for dissipative systems to accommodate L-infinity criteria (hard bound criteria). State feedback and measurement feedback synthesis procedures were derived for L-infinity-bounded robust control problems, and necessary and sufficient conditions were given using dynamic programming and information state techniques.

Nonlinear systems stability -- Input to State Stability (ISS) (with Matthew R. James, Dragan Nesic and Peter M. Dower)

Input-to-State Stability (ISS) is a natural generalization of internal stability for linear systems to nonlinear systems. Existing research work mainly focused on the equivalent definitions, ISS Lyapunov function characterizations, and ISS small gain theorems.

· Analysis (Automatica, 2005

)--- This work presented novel quantitative analysis results to characterise minimal ISS
gains and transient bounds. These characterisations naturally lead to computable necessary and sufficient conditions for ISS.

· Synthesis (IEEE Trans. on Automatic Control, 2005) --- This work presented a unified framework for the quantitative synthesis to achieve the input to state stability (ISS) property and other ISS-like properties with prescribed gain and transient bound. Quantitative results are very useful when applying the ISS small gain theorems. The results make a connection between ISS-like properties and optimization problems in the nonlinear dissipative systems theory (including L_2-gain analysis and nonlinear H-infinity theory).