Efficient Parallelized Simulation of Cyber-Physical Systems

Abstract

Advancements in accelerated physics simulations have greatly reduced training times for reinforcement learning policies, yet the conventional step-by-step agent-simulator interaction undermines simulation accuracy. In the real-world, interactions are asynchronous, with sensing, acting and processing happening simultaneously. Failing to capture this widens the sim2real gap and results in suboptimal real-world performance. In this paper, we address the challenges of simulating realistic asynchronicity and delays within parallelized simulations, crucial to bridging the sim2real gap in complex cyber-physical systems. Our approach efficiently parallelizes cyber-physical system simulations on accelerator hardware, including physics, sensors, actuators, processing components and their asynchronous interactions. We extend existing accelerated physics simulations with latency simulation capabilities by constructing a `supergraph' that encodes all data dependencies across parallelized simulation steps, ensuring accurate simulation. By finding the smallest supergraph, we minimize redundant computation. We validate our approach on two real-world systems and perform an extensive ablation, demonstrating superior performance compared to baseline methods.