Temporally Consistent Simulation of Robots and Their Controllers

Researchers simulate robot dynamics to optimize gains, trajectories, and controls and to validate proper robot operation. In this paper, we focus on this latter application, which allows roboticists to verify that robots do not damage themselves, the environments they are situated within, or humans. In current simulations, robot control code runs in lockstep with the dynamics integration. This design can result in code that appears viable in simulation but runs too slowly on physical systems. Addressing this problem requires overcoming significant challenges that arise due both to the speed of dynamic simulation running time (simulations may run 1/10 or 1/100 of real-time or slower) and to the variability of the running times (e.g., the speed of collision detection algorithms depends on pairwise object proximities). These difficulties imply that one must not only slow the control software but also scale controller running speeds dynamically. We describe the numerous architectural and OS-level technical challenges that we have overcome to yield temporally consistent simulation for modeling robots that use only real-time processes, and we show that our system is superior to the status quo using simulation-based experiments.