Semi-Supervised Trajectory-Feedback Controller Synthesis for Signal Temporal Logic Specifications
This work addresses the problem of real-time robot control with formal specifications for applications like self-driving vehicles, but it is incremental as it builds on prior STL optimization methods.
The paper tackles the challenge of incorporating spatio-temporal rules, expressed in Signal Temporal Logic, into robot control policies for real-time applications by proposing a semi-supervised controller synthesis technique that uses imitation-based regularization to induce human-like behaviors. The result shows that this approach leads to higher qualitative and quantitative performance and outperforms a state-of-the-art shooting method in both performance and computation time.
There are spatio-temporal rules that dictate how robots should operate in complex environments, e.g., road rules govern how (self-driving) vehicles should behave on the road. However, seamlessly incorporating such rules into a robot control policy remains challenging especially for real-time applications. In this work, given a desired spatio-temporal specification expressed in the Signal Temporal Logic (STL) language, we propose a semi-supervised controller synthesis technique that is attuned to human-like behaviors while satisfying desired STL specifications. Offline, we synthesize a trajectory-feedback neural network controller via an adversarial training scheme that summarizes past spatio-temporal behaviors when computing controls, and then online, we perform gradient steps to improve specification satisfaction. Central to the offline phase is an imitation-based regularization component that fosters better policy exploration and helps induce naturalistic human behaviors. Our experiments demonstrate that having imitation-based regularization leads to higher qualitative and quantitative performance compared to optimizing an STL objective only as done in prior work. We demonstrate the efficacy of our approach with an illustrative case study and show that our proposed controller outperforms a state-of-the-art shooting method in both performance and computation time.