Robots require the ability to autonomously and continuously react to unexpected online changes in the task definition and in the environment, especially those cohabited with humans. To react to these changes, the task, from the current state up to the finish, must instantly be reconsidered. This implies a prohibitive re-computation cost.
This paper proposes a modular control architecture based on Model Predictive Control, that offers a good compromise between optimally achieving the task and the required computation time, by only reconsidering the near future. This framework offers a generic way to formulate task-related objectives and constraints that dissociates the planning from the execution, which depends mainly on the robot dynamics.
The proposal exploits a linear formalization of the MPC in SE(3) to implement this architecture in a high-frequency closed-loop controller, achieving task re-planning at the control rate. The pertinence of the proposed control architecture is demonstrated using experiments with the Franka Emika robots in scenarios where the task to be achieved is modified on the fly.
The full version of the paper is open access and can be found in HAL database: manuscript