Section: New Results

Lexicographic Model Predictive Control for collision avoidance in dynamic environments

Participants : Pierre-Brice Wieber, Nestor Alonso Bohorquez Dorante, Alexander Sherikov, Dimitar Dimitrov.

Collision avoidance may not always be feasible in dynamic environments, when new obstacles can appear too late and move too fast with respect to the dynamic limitations of the system. A typical situation is with a biped robot walking in a compact and uncooperative crowd, with limited field of view. This year, we have investigated and compared 3 different relaxations of the collision avoidance constraint in this setting [33]. In the first case, collisions are accepted if the robot first comes to a stop, what corresponds to standard ISO norms for the safety of robots. In the second case, collisions are actively minimised by the robot, what gives significantly better results. In the third case, for the sake of completeness, the robot is allowed to fall in order to further avoid collisions. All three options were implemented with different formulations of lexicographic relaxation of the constraints in a standard MPC scheme for biped walking motion generation. This work raises important issues regarding safety norms for robots in human environments and how they are implemented.