Extending Open Dynamics Engine for the DARPA Virtual Robotics Challenge

Abstract

The DARPA Virtual Robotics Challenge (VRC) [1] was a cloud-based robotic simulation competition. Teams competed by writing control software for a humanoid robot to perform disaster response tasks in real-time simulation. Simulating the physics and sensors of a humanoid robot in real-time presented challenges related to the trade-off between simulation accuracy and computational time. The Projected Gauss-Seidel (PGS) iterative solver was chosen for its performance and robustness, but it lacks the accuracy and the fidelity required for reliable simulation of task-level behaviors. This paper presents the modeling decisions and algorithmic improvements made to the Open Dynamics Engine (ODE) physics solver that improved PGS accuracy and fidelity without sacrificing its real-time simulation performance in the VRC. These improvements allowed for stable simulation regardless of user input during the VRC, and supported reliable contact dynamics during VRC tasks without violating the near real-time requirement.