Learning feedback potential maps using large-scale optimal control

In recent work, an energy-aware control strategy was developed for robot manipulators, using offline task-based optimisation. The optimisation was cast into a time-varying impedance control framework, to yield an optimal open-loop control action. This open-loop action was supplemented by a low-gain task-free controller to reject unmodelled and external disturbances. The strategy uses a novel energy injection approach with tanks for increased safety, which provides a substantial benefit over
conventional energy tank approaches.

This work was extended to include a more comprehensive safety component, which uses energy-based detection of unintended interactions, and strategies for safe behaviour post-impact. Subsequently, this was experimentally verified on a Franka Emika Panda robot.

In the next step we would like to build on the existing work, in particularly in the following directions:

1) Better control over the task-based optimisation by using trajectory optimisation (optimal control) in CasaDI, to:

        - include pre-impact safety aspects,

        - increase insight into optimality, gradients, etc.

        - explore different frameworks for open-loop control

2) Further improvement on the safety aspects

3) Inclusion of (unknown) time-varying aspects such as interaction with
humans (e.g. in a tool handover task)

4) Experimental validation of these new developments