BERRY: Bit Error Robustness for Energy-Efficient Reinforcement Learning-Based Autonomous Systems

Abstract

Autonomous systems need to operate energy-efficiently due to their power and resource constraints. The low-voltage operation allows for further energy consumption reduction, however, causes memory bit-failures. This paper proposes a random bit-error-training technique, Berry, that improves the bit-error robustness in reinforcement learning-enabled autonomous systems. Berry leads to high energy savings in both compute-level operation and system-level quality-of-flight. Berry generalizes across operating voltages and accelerators, as demonstrated on bit errors from profiled SRAM arrays. Experiments on realistic virtual environments and real drone platform show that Berry significantly improves bit error robustness, saves flight energy, and increases the number of missions.