MIN-Fakultät
Fachbereich Informatik
TAMS

Biologically-Inspired Inverse Kinematics

Bachelor or Master thesis at group TAMS (also possible in cooperation with WTM).
Auch in deutscher Sprache möglich.

Motivation

Inverse kinematics constitues one of the most fundamental research problems in robotic systems and control of motion where numerous sophisticated algorithms have been presented during the last decades but no universal solution could yet be found. It is formulated by attempting to find a suitable configuration of joint variables satisfying a Cartesian objective and represents the opposite to forward kinematics which describes deriving a certain Cartesian result from a given consecutive joint variable configuration. While inverse kinematics has originally evolved from the discipline of robotics, significant importance was lately also gained in the field of character animation and interaction. Therefore, it is no more only related to industrial robot systems but also to applications settled in human-robot interaction, virtual reality and the movie and video game industry which represent an increasingly growing market and community.

The inverse kinematics problem is either solved by analytical or numerical approaches where the latter has been acknowledged a higher popularity in terms of universal applicability. However, typical problems still remain in suboptimal extrema configurations, comparatively high computation times or further requirements such as realistic motion and collision avoidance. During the last months, a novel biologically-inspired method based on genetic algorithms (GA) and particle swarm optimization (PSO) has been developed where experiments obtained 100% success rates with real-time capability on arbitrary kinematic structure. Accordingly, the mathematical problem for which no general solution exists is solved by evolutionary and collective principles as observed in natural phenomena.



The existing algorithm is implemented in C# using the Unity3D game engine and offers several interesting aspects for further investigations and improvements which reach from purely computational to algorithmic and scientific perspectives. The topics listed below are several thesis suggestions and offer much space for algorithmic research and experimentation but you are also free to come up with your own brilliant ideas! :-)

Topic 1: Parallelization


It was observed that >95% of the computation time was required for solving the mathematical forward kinematics equations. Accordingly, a reimplementation of the algorithm in C++/Python with an efficient parallel GPU or multi-core processing might achieve a significant computational speedup.

Research Target:

Requirements:


Topic 2: ROS Package


ROS (Robot Operating System) is a popular framework for robotics control. In this, several packages for solving the inverse kinematics problem already exist where a reimplementation and performance evaluation of the algorithm is of high interest. Also, there is a good chance that your open-source code will be used by many other people in robotics research! :-)

Research Target:

Requirements:


Topic 3: Parameter Study


For the algorithm, an extinction factor was introduced with intent to adaptively control the mutation of individuals in genetic algorithms. In this context, a study on the extinction factor might be considered where particular interest is in the caused loss by the adaptive control in contrast to model-specific optimized parameters.

Research Target:

Requirements:


Topic 4: Path Planning and/or Collision Avoidance


A common problem that is especially important in robotics and which directly constructs on the obtained solutions by inverse kinematics is given by path planning. Equivalently, the current implementation does not consider self-collision handling for the evolved joint variable configurations for but what is of crucial importance in real world applications.

Research Target:

Requirements:


Topic 5: Neural Learning


A very interesting study from a scientific perspective can be performed by integrating a neural learning of previously evolved solutions. Although artificial neural networks theirself show poor performance for solving the inverse kinematics problem as a whole, improvements can be imagined by feeding learned approximate configurations as guiding individuals into the evolutionary optimization. However, you are very welcome to develop your own ideas!

Research Target:

Requirements:


Contact