ADMM solver cuts computation for hierarchical robot control by up to 23.5%

Researchers led by Kai Pfeiffer have presented a solver based on the Alternating Direction Method of Multipliers (ADMM) for sparse hierarchical decision-making in robot control, such as choosing among many possible end-effector positions, in a paper posted on arXiv.

Existing solvers that use interior-point methods exploit the sparse, hierarchical structure of these problems but are hard to warm-start from one control step to the next, the authors write. Their ADMM formulation adds efficient substitutions for the variables that stand in for the l0-norm and evaluates linear algebra in an order that handles separable decision functions efficiently.

In large-scale kinematic and dynamic control, computation time fell by up to 23.5% compared with an interior-point method that exploits the problem structure, at low to moderate accuracy. Applying the approach to whole-body dynamic control for a walking humanoid showed that high-accuracy solutions are needed there, the authors report.