Core Concepts
The core message of this article is to provide a correct-by-design controller for an autonomous vehicle that can adjust its behavior based on the inferred intentions of opponent vehicles and pedestrians, while ensuring provable safety by restricting the probabilistic risk under a desired level.
Abstract
This paper presents a novel solution for intention-aware control of autonomous vehicles. The key highlights and insights are:
The authors use discrete-valued random variables to model the unknown intentions of opponent vehicles and pedestrians, capturing the complex epistemic uncertainty in their behaviors.
They formulate an intention-aware control problem with a belief-space specification, where the control objective is specified in the belief space using a Distribution Signal Temporal Logic (DSTL) formula.
To solve this challenging stochastic control problem, the authors leverage stochastic expansion techniques, particularly Polynomial Chaos Expansion (PCE), to convert the problem into a tractable deterministic control problem with an STL specification.
The solved intention-aware controller allows the autonomous vehicle to adjust its behaviors according to the inferred intentions of its opponents, ensuring provable safety by restricting the probabilistic risk under a desired level.
The authors validate the efficacy of their method through experimental studies on autonomous driving scenarios, including an overtaking case and an intersection case. The results show that the intention-aware controller can effectively avoid collisions with opponent vehicles and pedestrians, even in the presence of modeling uncertainties and linearization errors.
The proposed solution provides a novel framework for intention-aware control of autonomous systems with formal safety guarantees, addressing the limitations of existing planning-based methods that rely on precise agent models and struggle to ensure formal safety.