2019/09/11 by Carlos Mastalli, Rohan Budhiraja, Mastalli, Carlos +17 · 11 citations
Computer Science · Engineering · #Reinforcement Learning in Robotics #Adaptive Dynamic Programming Control #Fuel Cells and Related Materials
paper · pdf · doi:10.48550/arxiv.1909.04947
We introduce Crocoddyl (Contact RObot COntrol by Differential DYnamic\nLibrary), an open-source framework tailored for efficient multi-contact optimal\ncontrol. Crocoddyl efficiently computes the state trajectory and the control\npolicy for a given predefined sequence of contacts. Its efficiency is due to\nthe use of sparse analytical derivatives, exploitation of the problem\nstructure, and data sharing. It employs differential geometry to properly\ndescribe the state of any geometrical system, e.g. floating-base systems.\nAdditionally, we propose a novel optimal control algorithm called\nFeasibility-driven Differential Dynamic Programming (FDDP). Our method does not\nadd extra decision variables which often increases the computation time per\niteration due to factorization. FDDP shows a greater globalization strategy\ncompared to classical Differential Dynamic Programming (DDP) algorithms.\nConcretely, we propose two modifications to the classical DDP algorithm. First,\nthe backward pass accepts infeasible state-control trajectories. Second, the\nrollout keeps the gaps open during the early "exploratory" iterations (as\nexpected in multiple-shooting methods with only equality constraints). We\nshowcase the performance of our framework using different tasks. With our\nmethod, we can compute highly-dynamic maneuvers (e.g. jumping, front-flip)\nwithin few milliseconds.\n