Multicontact Whole Body Planning

This project focuses in our algorithm to generate dynamically feasible motions for a humanoid robot to traverse unstructured, confined spaces. As an example, we consider the case of traversing a knee knocker which requires simultaneously preserving balance and moving the robots’ limbs near their kinematic limits, especially for small- and medium-size robots.

The idea we propose is to treat the problem as a multi-particle planning problem with collision avoidance constraint. Then, we relax this nonlinear problem into a convex problem that provides kinematically feasible smooth trajectories for all of the planning frames (hands, knees, feet, torso). These paths are finally used to guide a full body trajectory optimization to verify the existence of joint plans that are dynamically feasible, i.e., respects joint angle limits, joint velocity limits, joint torque limits, and friction constraints.

A graphical video of our approach is shown below:

The details are shown in (Gonzalez & Sentis, 2024) and the code is available on GitHub.

Interactive animations of the resulting plans are available for the G1, ergoCub, and Valkyrie. To replay the animation, navigate the panel on the top-right corner by going to Open Controls -> Animations -> default ->play.