Effective Fault Tolerance for Robust Robotics under Radiation Exposure

The development of fault-tolerant autonomous robots for long-term deployment has been an area of active research for decades. Many researchers have focused on the tolerance to failures of sensors and actuators of various types of robots. In contrast to these failures, robots encounter transient faults on all levels including the bit level of microprocessors when operating in space or in nuclear facilities under high radiation. Additionally the shrinking nanometer technology of modern microprocessors and the supply voltage down-scaling drastically increase the risk of transient faults. Consequently even in nonhazardous environments modern processors encounter faults at an interval of days or hours. To cope with such soft errors, the VLSI community has developed numerous hardware hardening and low-level software protection schemes. Entire protection, however, makes processors usually slower, more expensive, and more energyconsuming. These standard techniques or an additional lead shielding are usually not desirable for autonomous robots. In this paper, we present novel techniques that bridge the gap between the low-level protection of microprocessors and high-level tolerance to sensor and actuator failures in robotics. Our methods enable a detailed analysis of applications under fault injection to identify critical parts of the combined hardware and software system. In our experiments, we demonstrate how our approach can be used to develop an efficient combination of selective hardening of processors and lightweight protection schemes of robotics software. The detection and isolation of actuator and sensor failures have been widely studied in robotics. There exist modelbased approaches that consider fault detection as a classification or regression problem and apply corresponding techniques [3, 4]. Furthermore the Bayesian approach has been popular to reason about discrete fault states during state estimation [8]. These techniques, however, apply severe model approximations or restrict the detection to faults that are specified in advance. Similar to our approach, Christensen et al. [2] apply fault injection during operation. As opposed to our method, they only inject numeric faults into the actuator and sensor data to generate data to train a classifier for fault detection. In contrast to the above mentioned approaches, Nagatani