Online Running Pattern Generation for Humanoid Robot With Direct Collocation of Reference-Tracking Dynamics

To generate dynamically feasible running motion for humanoid robots, one conventionally chooses between a simplified model and a full-body model. A simplified model can be calculated at high rates but ignores many of the robot's dynamic features, while a full-body model usually leads to time-consuming nonlinear programming. This article explores a powerful middle ground that takes the advantages of both worlds and proposes a running pattern generation method that can optimize the whole-body motion online. A three-body model generates the centroidal motion, including the center of mass position and the angular momentum. Then, we propose the so-called reference-tracking dynamics, where the collocated motion references in task space are projected to the joint space with serial quadratic programming. The separation of collocation and projection greatly reduces the scale of the problem. In addition, the two major motion references, centroidal and foot motion, can be handled separately to furtherly simplify the problem. Experiments on the humanoid robot platform BIT Humanoid Robot - Test (BHR-T) show that this method can generate fast and dynamically feasible running motion online, where running at 6 km/h is achieved.