CIMS - Robotics Research Seminars: Exploiti...

 CIMS - Robotics Research Seminars: Exploiting sensed contextual perception to resolve ambiguity in HRI (Expired)
Zhan Wang | Thursday May 03, 2012 | Research | Research

Date/Time: 12:30PM Monday May 07, 2012 - 1:30PM Monday May 07, 2012
Location: Room 7065, Level 7, Building 2
Download iCalendar

Dear All,

Please join the seminar by Mr. Dan Hordern from CIMS - Autonomous Robots Group. Dan is a current PhD student in the group. A light lunch will be provided.

Title: Exploiting sensed contextual perception to resolve ambiguity in HRI

Abstract:

Providing Robots with an understanding of how people act within an environment allows human behaviour to be incorporated into their decision making process. This human in the loop approach is important if robots are to be successfully introduced to applications requiring Human Robot Interaction (HRI).

Trajectory models have been used to add an element of social awareness in path planning and improve the tracking of people through short and long-term predictions inferred from the models. Whilst parameterised models (e.g. DBN's) have historically been utilised for these purposes, Gaussian Processes (GP) are a non-parametric model which have gained popularity in the literature for similar applications over recent years. Some beneficial properties of GP's are that they are continuous, do not need to have their structure explicitly defined, have been shown resistant to over-fitting and cope well with noisy data. Furthermore, the ability to integrate GP's into existing Bayesian filtering frameworks makes them suitable for this application. However, there are both general and application specific problems when using GP's, such as the computational cost as the size of the input data increases and the tendency for the GP trajectory model to generalise respectively. This research investigates the use of GP's to model trajectories and how to overcome some of these issues for this application. Furthermore, this research also examines the use of the learnt GP trajectory models with bearing sensors to improve their tracking and prediction capabilities.

Whilst understanding how people move within an environment is beneficial, it doesn't provide a very complete behavioural understanding of the environment. The auditory dimension provides essential information about the interactions of people which should be considered along with their motions. An example of this is in socially aware path planning, the robot should also consider whether it should move through quiet spaces or loud spaces. Therefore, this research also investigates how to practically map sound sources and incorporate this information into the robots decision making process.

The talk will also provide a practical introduction to Gaussian Processes Regression (GPR) and how GPR models can be adapted to the aforementioned application, including the general research challenges in applying GPR. It will also cover an introduction to acoustic beam forming and implementation considerations.

Date/time: 12.30 -- 1.30 pm, Monday, 7th of May, 2012

Location: Room 7065, Level 7, Building 2

Regards,

Zhan