School of
Information Technology and Electrical Engineering

Speaker: Benjamin Tam
Seminar Date: Wed, 22/01/2020 - 13:00
Venue: Michie Building (9) Room 221
Host: Prof Michael Bruenig

Seminar Type:  PhD Confirmation Seminar


Confined and complex subterranean environments present unknown hazards that robots can explore instead of humans for safety reasons. However, autonomous robots lack the capability to robustly select the best behaviour controller from a repertoire in a simple architecture, when faced with different situations. Additionally, robots could encounter paths blocked by or hidden behind movable obstacles, something current navigation planners assume are all static. However, if a multi-legged robot is able to detect if an obstacle is movable, manipulate the obstacle using its legs, and continue exploration; the robot can gather greater situational awareness for personnel outside.

To select the best behaviour for a given robot state, a probabilistic sequencing of behaviour controllers is used. This builds upon the concept of  Sequential Composition where a simple, funnel-like control architecture consists of controllers with outputs which prepare the system to subsequent controllers' inputs. Furthermore, a probabilistic framework fusing both exteroceptive and proprioceptive sensing of movable obstacles is used to update the costmap used for navigation planning. Using cameras and a 3D Lidar as exteroceptive sensors, images and a point cloud are used for object detection and classification. Through association of object classes to cost of manipulation, this additional information is added into the costmap for traditional navigation algorithms to find an efficient, potential path with manipulation. Proprioceptive information from feedback of wheeled/tracked robot velocity and motor torques are used to update the initial exteroceptive probability with actual cost. The object class costs can be updated with each new manipulation, learning the real world costs for the robot. Finally, these behaviours will be incorporated onto a legged robot platform, with leg manipulation of obstacles learnt through Reinforcement Learning and transferred onto the robot via system identification.

Initial testing and system development will utilise simulations. Preliminary performance of the system will be analysed on wheeled/tracked robot platforms before transferring onto a legged robot. Additionally, the aim is to deploy the system in real life scenarios at the DARPA Subterranean Challenge.


Benjamin Tam graduated with a BE(Hons) from the University of Queensland. He worked at CSIRO Data61’s Robotics and Autonomous Systems Group before deciding to pursue a doctorate degree. Currently working towards his PhD, he is supervised by Prof. Michael Bruenig, Dr. Navinda Kottege and Dr. Nicolas Hudson. His research interests include autonomy, legged robot locomotion, reinforcement learning and human–robot interaction.