Neural Simplex Architecture

Abstract

We present the Neural Simplex Architecture (NSA), a new approach to runtime assurance that provides safety guarantees for neural controllers (obtained e.g. using reinforcement learning) of autonomous and other complex systems without unduly sacrificing performance. NSA is inspired by the Simplex control architecture of Sha et al., but with some significant differences. In the traditional approach, the advanced controller (AC) is treated as a black box; when the decision module switches control to the baseline controller (BC), the BC remains in control forever. There is relatively little work on switching control back to the AC, and there are no techniques for correcting the AC's behavior after it generates a potentially unsafe control input that causes a failover to the BC. Our NSA addresses both of these limitations. NSA not only provides safety assurances in the presence of a possibly unsafe neural controller, but can also improve the safety of such a controller in an online setting via retraining, without overly degrading its performance. To demonstrate NSA's benefits, we have conducted several significant case studies in the continuous control domain. These include a targetseeking ground rover navigating an obstacle field, and a neural controller for an artificial pancreas system.

Towards safety monitoring of ML-based perception tasks of autonomous systems

“…For RQ4, there are promising alternatives for reacting when a detection is made. For example, using a modified simplex architecture [21] with two controllers: a high-performance and a verified-safe controller. In this architecture, the first controller uses ML, while the second one is based on a safety model.…”

Adaptive Simplex Architecture for Safe, Real-Time Robot Path Planning

“…Different versions of the SA have been used in robotic applications for autonomous vehicle [ 6 , 7 , 8 ] and drone flight path planning [ 9 ]. Where, in reference [ 9 ], the original SA is used, the other approaches use ML models as the HP planners.…”

“…In reference [ 9 ], a motion planning algorithm from the Open Motion Planning Library is used as the HP planner for a surveillance drone. In references [ 8 ] and [ 9 ], reinforcement learning-based HP planners are used for autonomous vehicle path planning. In reference [ 8 ], the so-called neural simplex architecture (NSA) is introduced, which is similar to the proposed ASA.…”

“…In references [ 8 ] and [ 9 ], reinforcement learning-based HP planners are used for autonomous vehicle path planning. In reference [ 8 ], the so-called neural simplex architecture (NSA) is introduced, which is similar to the proposed ASA. The NSA uses a deep neural network and reinforcement learning with rewards and penalties to train a ML model that is used as the HP planner in a rover navigation application with static obstacle avoidance.…”

“…ASA differs from NSA [ 8 ] in several significant ways. ASA uses a deterministic path planning algorithm (rather than a stochastic random walk heuristic) to generate training data for the supervised training of a feedforward neural network (or any other suitable ML method).…”