Learning the Kinematics of Tubular Continuum Robots

Model-based vs. Data-based Methods


Tubular continuum robots exhibit a curvilinear morphology which is compliant. Common robot modelling techniques to describe the relationship between actuator input in configuration space to the position and orientation in task space, and vice versa, are thus not applicable. Despite the simple actuation of component tubes, the resulting motion is characterised by a highly nonlinear behaviour due to the elastic interactions between the tubes. The current state of the art converged to approximating the curvilinear structure using continuum mechanics and formulating a set of nonlinear differential equations to solve for the quasi static shape of the robot numerically.

This model-­based approach sufficiently represents the underlying mechanics, but has limitations in terms of accuracy as well as efficiency. An inversion of the model can only be realised numerically by iteratively solving the forward model at high computational expense. High computational expense is also characterising state of the art methods in computational design optimization and motion planning, while not exploring the complete parameter space and thus leading to suboptimal results. Data­-based approaches, proposed here for the first time for tubular continuum robots, have the potential to overcome these limitations for tubular continuum robots. Deep learning can serve to discover unknown problem structures and to derive novel knowledge, which can then be used to improve and expand existing problem­ specific algorithms, but the merit is unexploited today.

The aim of this research programme is to leverage data­-based approaches and deep learning techniques for modelling, computational design, and motion planning for tubular continuum robots. The long term vision is to enable technology transfer of these techniques to real-­world applications of these robot, such as minimally invasive surgery, by focusing on the knowledge gaps for step change research. In the effort of achieving this, this research programme is planned around four objectives:

  • Define Data Representations for Tubular Continuum Robots
  • Investigate Deep Learning for Kinematic Modelling
  • Enable Task-­Optimal Robot Designs by Reinforcement Learning
  • Explore Learning­-based Motion Planning Techniques

Funding

  • NSERC Discovery Grant (April 1, 2019 – March 31, 2024)
  • NSERC Discovery Accelerator Supplements
  • NSERC Discovery Launch Supplements